release : v1.3.0

Also, fallback only if more temperatures are available and if we are
at least 3 seconds before the end of the audio

.github/workflows/bindings-go.yml

@@ -1,13 +1,18 @@
-name: Bindings Tests
+name: Bindings Tests (Go)
 on:
   push:
     paths:
       - bindings/go/**
+      - whisper.h
+  pull_request:
+    paths:
+      - bindings/go/**
+      - whisper.h
 
 jobs:
-    ubuntu-latest:
-      runs-on: ubuntu-latest
-      steps:
+  ubuntu-latest:
+    runs-on: ubuntu-latest
+    steps:
       - uses: actions/setup-go@v3
         with:
           go-version: '^1.19'

.github/workflows/bindings-ruby.yml

@@ -0,0 +1,22 @@
+name: Bindings Tests (Ruby)
+on:
+  push:
+    paths:
+      - bindings/ruby/**
+      - whisper.h
+  pull_request:
+    paths:
+      - bindings/ruby/**
+      - whisper.h
+
+jobs:
+  ubuntu-latest:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: ruby/setup-ruby@v1
+        with:
+          ruby-version: '3.0'
+      - uses: actions/checkout@v1
+      - run: |
+          cd bindings/ruby/ext
+          ruby extconf.rb && make

.github/workflows/build.yml

@@ -1,267 +1,308 @@
 name: CI
-on: [push]
+on: [push, pull_request]
 
 jobs:
-    ubuntu-latest:
-        runs-on: ubuntu-latest
+  ubuntu-latest:
+    runs-on: ubuntu-latest
 
-        steps:
-            - name: Clone
-              uses: actions/checkout@v1
+    steps:
+      - name: Clone
+        uses: actions/checkout@v1
 
-            - name: Dependencies
-              run: |
-                  sudo apt-get update
-                  sudo apt-get install build-essential
-                  sudo apt-get install libsdl2-dev
+      - name: Dependencies
+        run: |
+          sudo apt-get update
+          sudo apt-get install build-essential
+          sudo apt-get install libsdl2-dev
 
-            - name: Build
-              run: |
-                make
-                make stream
+      - name: Build
+        run: |
+          make
+          make stream
 
-    macOS-latest:
-        runs-on: macOS-latest
+  macOS-latest:
+    runs-on: macOS-latest
 
-        steps:
-            - name: Clone
-              uses: actions/checkout@v1
+    steps:
+      - name: Clone
+        uses: actions/checkout@v1
 
-            - name: Dependencies
-              run: |
-                  brew update
-                  brew install sdl2
+      - name: Dependencies
+        run: |
+          brew update
+          brew install sdl2
 
-            - name: Build
-              run: |
-                make
-                make stream
+      - name: Build
+        run: |
+          make
+          make stream
 
-    ubuntu-latest-gcc:
-        runs-on: ubuntu-latest
+  ubuntu-latest-gcc:
+    runs-on: ubuntu-latest
 
-        strategy:
-            matrix:
-                build: [Debug, Release]
+    strategy:
+      matrix:
+        build: [Debug, Release]
 
-        steps:
-            - name: Clone
-              uses: actions/checkout@v1
+    steps:
+      - name: Clone
+        uses: actions/checkout@v1
 
-            - name: Dependencies
-              run: |
-                  sudo apt-get update
-                  sudo apt-get install build-essential
-                  sudo apt-get install cmake
-                  sudo apt-get install libsdl2-dev
+      - name: Dependencies
+        run: |
+          sudo apt-get update
+          sudo apt-get install build-essential
+          sudo apt-get install cmake
+          sudo apt-get install libsdl2-dev
 
-            - name: Configure
-              run: cmake . -DWHISPER_SUPPORT_SDL2=ON -DCMAKE_BUILD_TYPE=${{ matrix.build }}
+      - name: Configure
+        run: cmake . -DWHISPER_SUPPORT_SDL2=ON -DCMAKE_BUILD_TYPE=${{ matrix.build }}
 
-            - name: Build
-              run: |
-                make
-                ctest -L gh --output-on-failure
+      - name: Build
+        run: |
+          make
+          ctest -L gh --output-on-failure
 
-    ubuntu-latest-clang:
-        runs-on: ubuntu-latest
+  ubuntu-latest-clang:
+    runs-on: ubuntu-latest
 
-        strategy:
-            matrix:
-                build: [Debug, Release]
+    strategy:
+      matrix:
+        build: [Debug, Release]
 
-        steps:
-            - name: Clone
-              uses: actions/checkout@v1
+    steps:
+      - name: Clone
+        uses: actions/checkout@v1
 
-            - name: Dependencies
-              run: |
-                  sudo apt-get update
-                  sudo apt-get install build-essential
-                  sudo apt-get install cmake
-                  sudo apt-get install libsdl2-dev
+      - name: Dependencies
+        run: |
+          sudo apt-get update
+          sudo apt-get install build-essential
+          sudo apt-get install cmake
+          sudo apt-get install libsdl2-dev
 
-            - name: Configure
-              run: cmake . -DWHISPER_SUPPORT_SDL2=ON -DCMAKE_BUILD_TYPE=${{ matrix.build }} -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_C_COMPILER=clang
+      - name: Configure
+        run: cmake . -DWHISPER_SUPPORT_SDL2=ON -DCMAKE_BUILD_TYPE=${{ matrix.build }} -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_C_COMPILER=clang
 
-            - name: Build
-              run: |
-                make
-                ctest -L gh --output-on-failure
+      - name: Build
+        run: |
+          make
+          ctest -L gh --output-on-failure
 
-    ubuntu-latest-gcc-sanitized:
-        runs-on: ubuntu-latest
+  ubuntu-latest-gcc-sanitized:
+    runs-on: ubuntu-latest
 
-        strategy:
-            matrix:
-                sanitizer: [ADDRESS, THREAD, UNDEFINED]
+    strategy:
+      matrix:
+        sanitizer: [ADDRESS, THREAD, UNDEFINED]
 
-        steps:
-            - name: Clone
-              uses: actions/checkout@v1
+    steps:
+      - name: Clone
+        uses: actions/checkout@v1
 
-            - name: Dependencies
-              run: |
-                  sudo apt-get update
-                  sudo apt-get install build-essential
-                  sudo apt-get install cmake
+      - name: Dependencies
+        run: |
+          sudo apt-get update
+          sudo apt-get install build-essential
+          sudo apt-get install cmake
 
-            - name: Configure
-              run: cmake . -DCMAKE_BUILD_TYPE=Debug -DWHISPER_SANITIZE_${{ matrix.sanitizer }}=ON
+      - name: Configure
+        run: cmake . -DCMAKE_BUILD_TYPE=Debug -DWHISPER_SANITIZE_${{ matrix.sanitizer }}=ON
 
-            - name: Build
-              run: |
-                make
-                ctest -L gh --output-on-failure
+      - name: Build
+        run: |
+          make
+          ctest -L gh --output-on-failure
 
-    windows:
-        runs-on: windows-latest
+  windows:
+    runs-on: windows-latest
 
-        strategy:
-            matrix:
-                build: [Release]
-                arch: [Win32, x64]
-                sdl2: [ON]
-                include:
-                  - arch: Win32
-                    s2arc: x86
-                  - arch: x64
-                    s2arc: x64
-                  - sdl2: ON
-                    s2ver: 2.26.0
+    strategy:
+      matrix:
+        build: [Release]
+        arch: [Win32, x64]
+        sdl2: [ON]
+        include:
+          - arch: Win32
+            s2arc: x86
+          - arch: x64
+            s2arc: x64
+          - sdl2: ON
+            s2ver: 2.26.0
 
-        steps:
-            - name: Clone
-              uses: actions/checkout@v1
+    steps:
+      - name: Clone
+        uses: actions/checkout@v1
 
-            - name: Add msbuild to PATH
-              uses: microsoft/setup-msbuild@v1
+      - name: Add msbuild to PATH
+        uses: microsoft/setup-msbuild@v1
 
-            - name: Fetch SDL2 and set SDL2_DIR
-              if: matrix.sdl2 == 'ON'
-              run: |
-                C:/msys64/usr/bin/wget.exe -qO sdl2.zip https://github.com/libsdl-org/SDL/releases/download/release-${{ matrix.s2ver }}/SDL2-devel-${{ matrix.s2ver }}-VC.zip
-                7z x sdl2.zip
-                echo "SDL2_DIR=$env:GITHUB_WORKSPACE/SDL2-${{ matrix.s2ver }}/cmake" >> $env:GITHUB_ENV
+      - name: Fetch SDL2 and set SDL2_DIR
+        if: matrix.sdl2 == 'ON'
+        run: |
+          C:/msys64/usr/bin/wget.exe -qO sdl2.zip https://github.com/libsdl-org/SDL/releases/download/release-${{ matrix.s2ver }}/SDL2-devel-${{ matrix.s2ver }}-VC.zip
+          7z x sdl2.zip
+          echo "SDL2_DIR=$env:GITHUB_WORKSPACE/SDL2-${{ matrix.s2ver }}/cmake" >> $env:GITHUB_ENV
 
-            - name: Configure
-              run: >
-                cmake -S . -B ./build -A ${{ matrix.arch }}
-                -DCMAKE_BUILD_TYPE=${{ matrix.build }}
-                -DWHISPER_SUPPORT_SDL2=${{ matrix.sdl2 }}
+      - name: Configure
+        run: >
+          cmake -S . -B ./build -A ${{ matrix.arch }}
+          -DCMAKE_BUILD_TYPE=${{ matrix.build }}
+          -DWHISPER_SUPPORT_SDL2=${{ matrix.sdl2 }}
 
-            - name: Build
-              run: |
-                cd ./build
-                msbuild ALL_BUILD.vcxproj -t:build -p:configuration=${{ matrix.build }} -p:platform=${{ matrix.arch }}
+      - name: Build
+        run: |
+          cd ./build
+          msbuild ALL_BUILD.vcxproj -t:build -p:configuration=${{ matrix.build }} -p:platform=${{ matrix.arch }}
 
-            - name: Copy SDL2.dll
-              if: matrix.sdl2 == 'ON'
-              run: copy "$env:SDL2_DIR/../lib/${{ matrix.s2arc }}/SDL2.dll" build/bin/${{ matrix.build }}
+      - name: Copy SDL2.dll
+        if: matrix.sdl2 == 'ON'
+        run: copy "$env:SDL2_DIR/../lib/${{ matrix.s2arc }}/SDL2.dll" build/bin/${{ matrix.build }}
 
-            - name: Upload binaries
-              if: matrix.sdl2 == 'ON'
-              uses: actions/upload-artifact@v1
-              with:
-                name: whisper-bin-${{ matrix.arch }}
-                path: build/bin/${{ matrix.build }}
+      - name: Upload binaries
+        if: matrix.sdl2 == 'ON'
+        uses: actions/upload-artifact@v1
+        with:
+          name: whisper-bin-${{ matrix.arch }}
+          path: build/bin/${{ matrix.build }}
 
-    windows-blas:
-        runs-on: windows-latest
+  windows-blas:
+    runs-on: windows-latest
 
-        strategy:
-            matrix:
-                build: [Release]
-                arch: [Win32, x64]
-                blas: [ON]
-                sdl2: [ON]
-                include:
-                  - arch: Win32
-                    obzip: https://github.com/xianyi/OpenBLAS/releases/download/v0.3.21/OpenBLAS-0.3.21-x86.zip
-                    s2arc: x86
-                  - arch: x64
-                    obzip: https://github.com/xianyi/OpenBLAS/releases/download/v0.3.21/OpenBLAS-0.3.21-x64.zip
-                    s2arc: x64
-                  - sdl2: ON
-                    s2ver: 2.26.0
+    strategy:
+      matrix:
+        build: [Release]
+        arch: [Win32, x64]
+        blas: [ON]
+        sdl2: [ON]
+        include:
+          - arch: Win32
+            obzip: https://github.com/xianyi/OpenBLAS/releases/download/v0.3.21/OpenBLAS-0.3.21-x86.zip
+            s2arc: x86
+          - arch: x64
+            obzip: https://github.com/xianyi/OpenBLAS/releases/download/v0.3.21/OpenBLAS-0.3.21-x64.zip
+            s2arc: x64
+          - sdl2: ON
+            s2ver: 2.26.0
 
-        steps:
-            - name: Clone
-              uses: actions/checkout@v1
+    steps:
+      - name: Clone
+        uses: actions/checkout@v1
 
-            - name: Add msbuild to PATH
-              uses: microsoft/setup-msbuild@v1
+      - name: Add msbuild to PATH
+        uses: microsoft/setup-msbuild@v1
 
-            - name: Fetch OpenBLAS
-              if: matrix.blas == 'ON'
-              run: |
-                C:/msys64/usr/bin/wget.exe -qO blas.zip ${{ matrix.obzip }}
-                7z x blas.zip -oblas -y
-                copy blas/include/cblas.h .
-                copy blas/include/openblas_config.h .
-                echo "blasdir=$env:GITHUB_WORKSPACE/blas" >> $env:GITHUB_ENV
+      - name: Fetch OpenBLAS
+        if: matrix.blas == 'ON'
+        run: |
+          C:/msys64/usr/bin/wget.exe -qO blas.zip ${{ matrix.obzip }}
+          7z x blas.zip -oblas -y
+          copy blas/include/cblas.h .
+          copy blas/include/openblas_config.h .
+          echo "blasdir=$env:GITHUB_WORKSPACE/blas" >> $env:GITHUB_ENV
 
-            - name: Fetch SDL2 and set SDL2_DIR
-              if: matrix.sdl2 == 'ON'
-              run: |
-                C:/msys64/usr/bin/wget.exe -qO sdl2.zip https://github.com/libsdl-org/SDL/releases/download/release-${{ matrix.s2ver }}/SDL2-devel-${{ matrix.s2ver }}-VC.zip
-                7z x sdl2.zip
-                echo "SDL2_DIR=$env:GITHUB_WORKSPACE/SDL2-${{ matrix.s2ver }}/cmake" >> $env:GITHUB_ENV
+      - name: Fetch SDL2 and set SDL2_DIR
+        if: matrix.sdl2 == 'ON'
+        run: |
+          C:/msys64/usr/bin/wget.exe -qO sdl2.zip https://github.com/libsdl-org/SDL/releases/download/release-${{ matrix.s2ver }}/SDL2-devel-${{ matrix.s2ver }}-VC.zip
+          7z x sdl2.zip
+          echo "SDL2_DIR=$env:GITHUB_WORKSPACE/SDL2-${{ matrix.s2ver }}/cmake" >> $env:GITHUB_ENV
 
-            - name: Configure
-              run: >
-                cmake -S . -B ./build -A ${{ matrix.arch }}
-                -DCMAKE_BUILD_TYPE=${{ matrix.build }}
-                -DWHISPER_SUPPORT_OPENBLAS=${{ matrix.blas }}
-                -DCMAKE_LIBRARY_PATH="$env:blasdir/lib"
-                -DWHISPER_SUPPORT_SDL2=${{ matrix.sdl2 }}
+      - name: Configure
+        run: >
+          cmake -S . -B ./build -A ${{ matrix.arch }}
+          -DCMAKE_BUILD_TYPE=${{ matrix.build }}
+          -DWHISPER_SUPPORT_OPENBLAS=${{ matrix.blas }}
+          -DCMAKE_LIBRARY_PATH="$env:blasdir/lib"
+          -DWHISPER_SUPPORT_SDL2=${{ matrix.sdl2 }}
 
-            - name: Build
-              run: |
-                cd ./build
-                msbuild ALL_BUILD.vcxproj -t:build -p:configuration=${{ matrix.build }} -p:platform=${{ matrix.arch }}
+      - name: Build
+        run: |
+          cd ./build
+          msbuild ALL_BUILD.vcxproj -t:build -p:configuration=${{ matrix.build }} -p:platform=${{ matrix.arch }}
 
-            - name: Copy libopenblas.dll
-              if: matrix.blas == 'ON'
-              run: copy "$env:blasdir/bin/libopenblas.dll" build/bin/${{ matrix.build }}
+      - name: Copy libopenblas.dll
+        if: matrix.blas == 'ON'
+        run: copy "$env:blasdir/bin/libopenblas.dll" build/bin/${{ matrix.build }}
 
-            - name: Copy SDL2.dll
-              if: matrix.sdl2 == 'ON'
-              run: copy "$env:SDL2_DIR/../lib/${{ matrix.s2arc }}/SDL2.dll" build/bin/${{ matrix.build }}
+      - name: Copy SDL2.dll
+        if: matrix.sdl2 == 'ON'
+        run: copy "$env:SDL2_DIR/../lib/${{ matrix.s2arc }}/SDL2.dll" build/bin/${{ matrix.build }}
 
-            - name: Upload binaries
-              if: matrix.blas == 'ON' && matrix.sdl2 == 'ON'
-              uses: actions/upload-artifact@v1
-              with:
-                name: whisper-blas-bin-${{ matrix.arch }}
-                path: build/bin/${{ matrix.build }}
+      - name: Upload binaries
+        if: matrix.blas == 'ON' && matrix.sdl2 == 'ON'
+        uses: actions/upload-artifact@v1
+        with:
+          name: whisper-blas-bin-${{ matrix.arch }}
+          path: build/bin/${{ matrix.build }}
 
-    emscripten:
-        runs-on: ubuntu-latest
+  emscripten:
+    runs-on: ubuntu-latest
 
-        strategy:
-            matrix:
-                build: [Release]
+    strategy:
+      matrix:
+        build: [Release]
 
-        steps:
-            - name: Clone
-              uses: actions/checkout@v1
+    steps:
+      - name: Clone
+        uses: actions/checkout@v1
 
-            - name: Dependencies
-              run: |
-                wget -q https://github.com/emscripten-core/emsdk/archive/master.tar.gz
-                tar -xvf master.tar.gz
-                emsdk-master/emsdk update
-                emsdk-master/emsdk install latest
-                emsdk-master/emsdk activate latest
+      - name: Dependencies
+        run: |
+          wget -q https://github.com/emscripten-core/emsdk/archive/master.tar.gz
+          tar -xvf master.tar.gz
+          emsdk-master/emsdk update
+          emsdk-master/emsdk install latest
+          emsdk-master/emsdk activate latest
 
-            - name: Configure
-              run: echo "tmp"
+      - name: Configure
+        run: echo "tmp"
 
-            - name: Build
-              run: |
-                pushd emsdk-master
-                source ./emsdk_env.sh
-                popd
-                emcmake cmake . -DCMAKE_BUILD_TYPE=${{ matrix.build }}
-                make
+      - name: Build
+        run: |
+          pushd emsdk-master
+          source ./emsdk_env.sh
+          popd
+          emcmake cmake . -DCMAKE_BUILD_TYPE=${{ matrix.build }}
+          make
+
+  ios:
+    runs-on: macos-latest
+
+    strategy:
+      matrix:
+        build: [Release]
+
+    steps:
+      - name: Clone
+        uses: actions/checkout@v1
+
+      - name: Configure
+        run: cp models/for-tests-ggml-base.en.bin models/ggml-base.en.bin
+
+      - name: Build objc example
+        run: xcodebuild -project examples/whisper.objc/whisper.objc.xcodeproj -scheme whisper.objc -configuration ${{ matrix.build }} -sdk iphonesimulator build
+      
+      - name: Build swiftui example
+        run: xcodebuild -project examples/whisper.swiftui/whisper.swiftui.xcodeproj -scheme WhisperCppDemo -configuration ${{ matrix.build }} -sdk iphonesimulator build
+
+  android:
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Clone
+        uses: actions/checkout@v1
+
+      - name: Install Java
+        uses: actions/setup-java@v3
+        with:
+          distribution: zulu
+          java-version: 17
+      
+      - name: Setup Android SDK
+        uses: android-actions/setup-android@v2
+
+      - name: Build
+        run: |
+          cd examples/whisper.android
+          ./gradlew assembleRelease --no-daemon

.github/workflows/examples.yml

@@ -0,0 +1,48 @@
+name: Examples Tests
+on:
+  push:
+    paths:
+      - examples/addon.node/**
+      - whisper.h
+  pull_request:
+    paths:
+      - examples/addon.node/**
+      - whisper.h
+
+jobs:
+  addon_node-ubuntu-latest:
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        node-version: [ 16.x, 18.x ]
+    steps:
+      - name: Clone
+        uses: actions/checkout@v1
+
+      - name: Dependencies
+        run: |
+          sudo apt-get update
+          sudo apt-get install build-essential
+          sudo apt-get install cmake
+          sudo apt-get install libsdl2-dev
+
+      - name: Use Node.js ${{ matrix.node-version }}
+        uses: actions/setup-node@v1
+        with:
+          node-version: ${{ matrix.node-version }}
+          cache: 'npm'
+
+      - name: Install package.json dependencies
+        working-directory: ./examples/addon.node
+        run: npm install
+
+      - name: Compile addon.node
+        run: npx cmake-js compile -T whisper-addon -B Release
+
+      - name: Download test model
+        run: |
+          bash ./models/download-ggml-model.sh base.en
+      - name: Test
+        run: |
+          cd examples/addon.node
+          npm run test

.gitignore

@@ -1,5 +1,8 @@
 *.o
+*.a
 .cache/
+.coreml/
+.test/
 .vs/
 .vscode/
 .DS_Store
@@ -9,6 +12,7 @@ build-em/
 build-debug/
 build-release/
 build-static/
+build-no-accel/
 build-sanitize-addr/
 build-sanitize-thread/
 
@@ -16,8 +20,10 @@ build-sanitize-thread/
 /stream
 /command
 /talk
+/talk-llama
 /bench
 
+arm_neon.h
 sync.sh
 libwhisper.a
 libwhisper.so
@@ -29,3 +35,7 @@ examples/whisper.objc/whisper.objc.xcodeproj/xcuserdata/
 examples/whisper.objc/whisper.objc.xcodeproj/project.xcworkspace/xcuserdata
 
 extra/bench-gg.txt
+
+models/*.mlmodel
+models/*.mlmodelc
+models/*.mlpackage

CMakeLists.txt

@@ -1,6 +1,10 @@
 cmake_minimum_required (VERSION 3.0)
 
-project(whisper.cpp VERSION 1.1.0)
+project(whisper.cpp VERSION 1.3.0)
+
+if ("${CMAKE_CXX_COMPILER_ID}" STREQUAL "MSVC")
+    add_compile_options(/utf-8)
+endif ()
 
 # Add path to modules
 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
@@ -54,6 +58,8 @@ if (APPLE)
     option(WHISPER_NO_AVX              "whisper: disable AVX" OFF)
     option(WHISPER_NO_AVX2             "whisper: disable AVX2" OFF)
     option(WHISPER_NO_FMA              "whisper: disable FMA" OFF)
+
+    option(WHISPER_COREML              "whisper: enable Core ML framework" OFF)
 else()
     option(WHISPER_SUPPORT_OPENBLAS    "whisper: support for OpenBLAS" OFF)
 endif()
@@ -86,16 +92,33 @@ endif()
 
 find_package(Threads REQUIRED)
 
-# on APPLE - include Accelerate framework
-if (APPLE AND NOT WHISPER_NO_ACCELERATE)
-    find_library(ACCELERATE_FRAMEWORK Accelerate)
-    if (ACCELERATE_FRAMEWORK)
-        message(STATUS "Accelerate framework found")
+# on APPLE
+if (APPLE)
+    # include Accelerate framework
+    if (NOT WHISPER_NO_ACCELERATE)
+        find_library(ACCELERATE_FRAMEWORK Accelerate)
 
-        set(WHISPER_EXTRA_LIBS  ${WHISPER_EXTRA_LIBS}  ${ACCELERATE_FRAMEWORK})
-        set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_ACCELERATE)
-    else()
-        message(WARNING "Accelerate framework not found")
+        if (ACCELERATE_FRAMEWORK)
+            message(STATUS "Accelerate framework found")
+
+            set(WHISPER_EXTRA_LIBS  ${WHISPER_EXTRA_LIBS}  ${ACCELERATE_FRAMEWORK})
+            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_ACCELERATE)
+        else()
+            message(WARNING "Accelerate framework not found")
+        endif()
+    endif()
+
+    if (WHISPER_COREML)
+        find_library(FOUNDATION_FRAMEWORK Foundation)
+        find_library(COREML_FRAMEWORK CoreML)
+
+        if (COREML_FRAMEWORK)
+            message(STATUS "CoreML framework found")
+
+            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DWHISPER_USE_COREML)
+        else()
+            message(WARNING "CoreML framework not found")
+        endif()
     endif()
 endif()
 
@@ -172,7 +195,9 @@ else()
             if(NOT WHISPER_NO_FMA)
                 set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mfma")
             endif()
-            set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mf16c")
+            if(NOT WHISPER_NO_F16C)
+                set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mf16c")
+            endif()
         endif()
     endif()
 endif()
@@ -181,6 +206,33 @@ if (WHISPER_PERF)
     set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_PERF)
 endif()
 
+#
+# whisper.coreml - Core ML support
+#
+
+if (WHISPER_COREML)
+    set(TARGET whisper.coreml)
+
+    add_library(${TARGET}
+        coreml/whisper-encoder.h
+        coreml/whisper-encoder.mm
+        coreml/whisper-encoder-impl.h
+        coreml/whisper-encoder-impl.m
+        )
+
+    include(DefaultTargetOptions)
+
+    target_include_directories(${TARGET} PUBLIC
+        .
+        )
+
+    target_link_libraries(${TARGET} PRIVATE ${FOUNDATION_FRAMEWORK} ${COREML_FRAMEWORK})
+
+    set_target_properties(${TARGET} PROPERTIES
+        COMPILE_FLAGS "-fobjc-arc"
+        )
+endif()
+
 #
 # whisper - this is the main library of the project
 #
@@ -200,6 +252,10 @@ target_include_directories(${TARGET} PUBLIC
     .
     )
 
+if (WHISPER_COREML)
+    target_link_libraries(${TARGET} PRIVATE whisper.coreml)
+endif()
+
 if (MSVC)
     target_link_libraries(${TARGET} PRIVATE ${WHISPER_EXTRA_LIBS} ${CMAKE_THREAD_LIBS_INIT})
 
@@ -226,10 +282,13 @@ target_compile_definitions(${TARGET} PUBLIC
     ${WHISPER_EXTRA_FLAGS}
     )
 
+set_target_properties(${TARGET} PROPERTIES PUBLIC_HEADER "whisper.h")
+
 install(TARGETS ${TARGET}
     LIBRARY DESTINATION lib
     ARCHIVE DESTINATION lib/static
     RUNTIME DESTINATION bin
+    PUBLIC_HEADER DESTINATION include
     )
 
 #
@@ -242,7 +301,7 @@ add_subdirectory(bindings)
 # programs, examples and tests
 #
 
-if (WHISPER_BUILD_TESTS)
+if (WHISPER_BUILD_TESTS AND NOT CMAKE_JS_VERSION)
     enable_testing()
     add_subdirectory(tests)
 endif ()

LICENSE

@@ -1,6 +1,6 @@
 MIT License
 
-Copyright (c) 2022 Georgi Gerganov
+Copyright (c) 2023 Georgi Gerganov
 
 Permission is hereby granted, free of charge, to any person obtaining a copy
 of this software and associated documentation files (the "Software"), to deal

Makefile

@@ -30,10 +30,16 @@ endif
 # Compile flags
 #
 
-CFLAGS   = -I.              -O3 -std=c11   -fPIC
-CXXFLAGS = -I. -I./examples -O3 -std=c++11 -fPIC
+CFLAGS   = -I.              -O3 -DNDEBUG -std=c11   -fPIC
+CXXFLAGS = -I. -I./examples -O3 -DNDEBUG -std=c++11 -fPIC
 LDFLAGS  =
 
+# ref: https://github.com/ggerganov/whisper.cpp/issues/37
+ifneq ($(wildcard /usr/include/musl/*),)
+	CFLAGS += -D_POSIX_SOURCE -D_GNU_SOURCE
+	CXXFLAGS += -D_POSIX_SOURCE -D_GNU_SOURCE
+endif
+
 # OS specific
 # TODO: support Windows
 ifeq ($(UNAME_S),Linux)
@@ -71,10 +77,6 @@ ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686))
 			CFLAGS += -mavx2
 		endif
 	else ifeq ($(UNAME_S),Linux)
-		AVX1_M := $(shell grep "avx " /proc/cpuinfo)
-		ifneq (,$(findstring avx,$(AVX1_M)))
-			CFLAGS += -mavx
-		endif
 		AVX2_M := $(shell grep "avx2 " /proc/cpuinfo)
 		ifneq (,$(findstring avx2,$(AVX2_M)))
 			CFLAGS += -mavx2
@@ -86,16 +88,17 @@ ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686))
 		F16C_M := $(shell grep "f16c " /proc/cpuinfo)
 		ifneq (,$(findstring f16c,$(F16C_M)))
 			CFLAGS += -mf16c
+
+			AVX1_M := $(shell grep "avx " /proc/cpuinfo)
+			ifneq (,$(findstring avx,$(AVX1_M)))
+				CFLAGS += -mavx
+			endif
 		endif
 		SSE3_M := $(shell grep "sse3 " /proc/cpuinfo)
 		ifneq (,$(findstring sse3,$(SSE3_M)))
 			CFLAGS += -msse3
 		endif
 	else ifeq ($(UNAME_S),Haiku)
-		AVX1_M := $(shell sysinfo -cpu | grep "AVX ")
-		ifneq (,$(findstring avx,$(AVX1_M)))
-			CFLAGS += -mavx
-		endif
 		AVX2_M := $(shell sysinfo -cpu | grep "AVX2 ")
 		ifneq (,$(findstring avx2,$(AVX2_M)))
 			CFLAGS += -mavx2
@@ -107,6 +110,11 @@ ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686))
 		F16C_M := $(shell sysinfo -cpu | grep "F16C ")
 		ifneq (,$(findstring f16c,$(F16C_M)))
 			CFLAGS += -mf16c
+
+			AVX1_M := $(shell sysinfo -cpu | grep "AVX ")
+			ifneq (,$(findstring avx,$(AVX1_M)))
+				CFLAGS += -mavx
+			endif
 		endif
 	else
 		CFLAGS += -mfma -mf16c -mavx -mavx2
@@ -115,11 +123,15 @@ endif
 ifeq ($(UNAME_M),amd64)
 	CFLAGS += -mavx -mavx2 -mfma -mf16c
 endif
-ifeq ($(UNAME_M),ppc64le)
+ifneq ($(filter ppc64%,$(UNAME_M)),)
 	POWER9_M := $(shell grep "POWER9" /proc/cpuinfo)
 	ifneq (,$(findstring POWER9,$(POWER9_M)))
 		CFLAGS += -mpower9-vector
 	endif
+	# Require c++23's std::byteswap for big-endian support.
+	ifeq ($(UNAME_M),ppc64)
+		CXXFLAGS += -std=c++23 -DGGML_BIG_ENDIAN
+	endif
 endif
 ifndef WHISPER_NO_ACCELERATE
 	# Mac M1 - include Accelerate framework
@@ -128,23 +140,32 @@ ifndef WHISPER_NO_ACCELERATE
 		LDFLAGS += -framework Accelerate
 	endif
 endif
+ifdef WHISPER_COREML
+	CXXFLAGS += -DWHISPER_USE_COREML
+	LDFLAGS  += -framework Foundation -framework CoreML
+endif
 ifdef WHISPER_OPENBLAS
 	CFLAGS  += -DGGML_USE_OPENBLAS -I/usr/local/include/openblas
 	LDFLAGS += -lopenblas
 endif
 ifdef WHISPER_GPROF
-	CFLAGS  += -pg
-	CXXFLAGS  += -pg
+	CFLAGS   += -pg
+	CXXFLAGS += -pg
 endif
 ifneq ($(filter aarch64%,$(UNAME_M)),)
+	CFLAGS += -mcpu=native
+	CXXFLAGS += -mcpu=native
 endif
 ifneq ($(filter armv6%,$(UNAME_M)),)
-	# Raspberry Pi 1, 2, 3
-	CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access
+	# 32-bit Raspberry Pi 1, 2, 3
+	CFLAGS += -mfpu=neon -mfp16-format=ieee -mno-unaligned-access
 endif
 ifneq ($(filter armv7%,$(UNAME_M)),)
-	# Raspberry Pi 4
-	CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations
+	# 32-bit ARM, for example on Armbian or possibly raspbian
+	CFLAGS += -mfpu=neon -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations
+
+	# 64-bit ARM, use these (TODO: auto-detect 64-bit)
+	# CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations
 endif
 ifneq ($(filter armv8%,$(UNAME_M)),)
 	# Raspberry Pi 4
@@ -166,7 +187,7 @@ $(info I CC:       $(CCV))
 $(info I CXX:      $(CXXV))
 $(info )
 
-default: main
+default: main bench
 
 #
 # Build library
@@ -175,17 +196,29 @@ default: main
 ggml.o: ggml.c ggml.h
 	$(CC)  $(CFLAGS)   -c ggml.c -o ggml.o
 
-whisper.o: whisper.cpp whisper.h
+whisper.o: whisper.cpp whisper.h ggml.h
 	$(CXX) $(CXXFLAGS) -c whisper.cpp -o whisper.o
 
-libwhisper.a: ggml.o whisper.o
-	$(AR) rcs libwhisper.a ggml.o whisper.o
+ifndef WHISPER_COREML
+WHISPER_OBJ = whisper.o
+else
+whisper-encoder.o: coreml/whisper-encoder.mm coreml/whisper-encoder.h
+	$(CXX) -O3 -I . -c coreml/whisper-encoder.mm -o whisper-encoder.o
 
-libwhisper.so: ggml.o whisper.o
-	$(CXX) $(CXXFLAGS) -shared -o libwhisper.so ggml.o whisper.o $(LDFLAGS)
+whisper-encoder-impl.o: coreml/whisper-encoder-impl.m coreml/whisper-encoder-impl.h
+	$(CXX) -O3 -I . -fobjc-arc -c coreml/whisper-encoder-impl.m -o whisper-encoder-impl.o
+
+WHISPER_OBJ = whisper.o whisper-encoder.o whisper-encoder-impl.o
+endif
+
+libwhisper.a: ggml.o $(WHISPER_OBJ)
+	$(AR) rcs libwhisper.a ggml.o $(WHISPER_OBJ)
+
+libwhisper.so: ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) -shared -o libwhisper.so ggml.o $(WHISPER_OBJ) $(LDFLAGS)
 
 clean:
-	rm -f *.o main stream command talk bench libwhisper.a libwhisper.so
+	rm -f *.o main stream command talk talk-llama bench libwhisper.a libwhisper.so
 
 #
 # Examples
@@ -193,21 +226,27 @@ clean:
 
 CC_SDL=`sdl2-config --cflags --libs`
 
-main: examples/main/main.cpp ggml.o whisper.o
-	$(CXX) $(CXXFLAGS) examples/main/main.cpp ggml.o whisper.o -o main $(LDFLAGS)
+SRC_COMMON = examples/common.cpp
+SRC_COMMON_SDL = examples/common-sdl.cpp
+
+main: examples/main/main.cpp $(SRC_COMMON) ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/main/main.cpp $(SRC_COMMON) ggml.o $(WHISPER_OBJ) -o main $(LDFLAGS)
 	./main -h
 
-stream: examples/stream/stream.cpp ggml.o whisper.o
-	$(CXX) $(CXXFLAGS) examples/stream/stream.cpp ggml.o whisper.o -o stream $(CC_SDL) $(LDFLAGS)
+bench: examples/bench/bench.cpp ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/bench/bench.cpp ggml.o $(WHISPER_OBJ) -o bench $(LDFLAGS)
 
-command: examples/command/command.cpp ggml.o whisper.o
-	$(CXX) $(CXXFLAGS) examples/command/command.cpp ggml.o whisper.o -o command $(CC_SDL) $(LDFLAGS)
+stream: examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ) -o stream $(CC_SDL) $(LDFLAGS)
 
-talk: examples/talk/talk.cpp  examples/talk/gpt-2.cpp ggml.o whisper.o
-	$(CXX) $(CXXFLAGS) examples/talk/talk.cpp examples/talk/gpt-2.cpp ggml.o whisper.o -o talk $(CC_SDL) $(LDFLAGS)
+command: examples/command/command.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/command/command.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ) -o command $(CC_SDL) $(LDFLAGS)
 
-bench: examples/bench/bench.cpp ggml.o whisper.o
-	$(CXX) $(CXXFLAGS) examples/bench/bench.cpp ggml.o whisper.o -o bench $(LDFLAGS)
+talk: examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ) -o talk $(CC_SDL) $(LDFLAGS)
+
+talk-llama: examples/talk-llama/talk-llama.cpp examples/talk-llama/llama.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/talk-llama/talk-llama.cpp examples/talk-llama/llama.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ) -o talk-llama $(CC_SDL) $(LDFLAGS)
 
 #
 # Audio samples

README.md

@@ -4,16 +4,16 @@
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
 [![npm](https://img.shields.io/npm/v/whisper.cpp.svg)](https://www.npmjs.com/package/whisper.cpp/)
 
-Stable: [1.0.4](https://github.com/ggerganov/whisper.cpp/releases/tag/1.0.4) / Beta: [1.1.0](https://github.com/ggerganov/whisper.cpp/releases/tag/1.1.0) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)
+Beta: [v1.3.0](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.3.0) / Stable: [v1.2.1](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.2.1) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)
 
 High-performance inference of [OpenAI's Whisper](https://github.com/openai/whisper) automatic speech recognition (ASR) model:
 
 - Plain C/C++ implementation without dependencies
-- Apple silicon first-class citizen - optimized via Arm Neon and Accelerate framework
+- Apple silicon first-class citizen - optimized via ARM NEON, Accelerate framework and [Core ML](https://github.com/ggerganov/whisper.cpp#core-ml-support)
 - AVX intrinsics support for x86 architectures
 - VSX intrinsics support for POWER architectures
 - Mixed F16 / F32 precision
-- Low memory usage (Flash Attention + Flash Forward)
+- Low memory usage (Flash Attention)
 - Zero memory allocations at runtime
 - Runs on the CPU
 - [C-style API](https://github.com/ggerganov/whisper.cpp/blob/master/whisper.h)
@@ -58,7 +58,9 @@ the Accelerate framework utilizes the special-purpose AMX coprocessor available
 
 ## Quick start
 
-First, download one of the Whisper models converted in [ggml format](models). For example:
+First clone the repository.
+
+Then, download one of the Whisper models converted in [ggml format](models). For example:
 
 ```bash
 bash ./models/download-ggml-model.sh base.en
@@ -89,35 +91,37 @@ c++ -I. -I./examples -O3 -std=c++11 -pthread examples/main/main.cpp whisper.o gg
 usage: ./main [options] file0.wav file1.wav ...
 
 options:
-  -h,       --help            [default] show this help message and exit
-  -t N,     --threads N       [4      ] number of threads to use during computation
-  -p N,     --processors N    [1      ] number of processors to use during computation
-  -ot N,    --offset-t N      [0      ] time offset in milliseconds
-  -on N,    --offset-n N      [0      ] segment index offset
-  -d  N,    --duration N      [0      ] duration of audio to process in milliseconds
-  -mc N,    --max-context N   [-1     ] maximum number of text context tokens to store
-  -ml N,    --max-len N       [0      ] maximum segment length in characters
-  -bo N,    --best-of N       [5      ] number of best candidates to keep
-  -bs N,    --beam-size N     [-1     ] beam size for beam search
-  -wt N,    --word-thold N    [0.01   ] word timestamp probability threshold
-  -et N,    --entropy-thold N [2.40   ] entropy threshold for decoder fail
-  -lpt N,   --logprob-thold N [-1.00  ] log probability threshold for decoder fail
-  -su,      --speed-up        [false  ] speed up audio by x2 (reduced accuracy)
-  -tr,      --translate       [false  ] translate from source language to english
-  -di,      --diarize         [false  ] stereo audio diarization
-  -otxt,    --output-txt      [false  ] output result in a text file
-  -ovtt,    --output-vtt      [false  ] output result in a vtt file
-  -osrt,    --output-srt      [false  ] output result in a srt file
-  -owts,    --output-words    [false  ] output script for generating karaoke video
-  -ocsv,    --output-csv      [false  ] output result in a CSV file
-  -ps,      --print-special   [false  ] print special tokens
-  -pc,      --print-colors    [false  ] print colors
-  -pp,      --print-progress  [false  ] print progress
-  -nt,      --no-timestamps   [true   ] do not print timestamps
-  -l LANG,  --language LANG   [en     ] spoken language ('auto' for auto-detect)
-            --prompt PROMPT   [       ] initial prompt
-  -m FNAME, --model FNAME     [models/ggml-base.en.bin] model path
-  -f FNAME, --file FNAME      [       ] input WAV file path
+  -h,        --help              [default] show this help message and exit
+  -t N,      --threads N         [4      ] number of threads to use during computation
+  -p N,      --processors N      [1      ] number of processors to use during computation
+  -ot N,     --offset-t N        [0      ] time offset in milliseconds
+  -on N,     --offset-n N        [0      ] segment index offset
+  -d  N,     --duration N        [0      ] duration of audio to process in milliseconds
+  -mc N,     --max-context N     [-1     ] maximum number of text context tokens to store
+  -ml N,     --max-len N         [0      ] maximum segment length in characters
+  -bo N,     --best-of N         [5      ] number of best candidates to keep
+  -bs N,     --beam-size N       [-1     ] beam size for beam search
+  -wt N,     --word-thold N      [0.01   ] word timestamp probability threshold
+  -et N,     --entropy-thold N   [2.40   ] entropy threshold for decoder fail
+  -lpt N,    --logprob-thold N   [-1.00  ] log probability threshold for decoder fail
+  -su,       --speed-up          [false  ] speed up audio by x2 (reduced accuracy)
+  -tr,       --translate         [false  ] translate from source language to english
+  -di,       --diarize           [false  ] stereo audio diarization
+  -nf,       --no-fallback       [false  ] do not use temperature fallback while decoding
+  -otxt,     --output-txt        [false  ] output result in a text file
+  -ovtt,     --output-vtt        [false  ] output result in a vtt file
+  -osrt,     --output-srt        [false  ] output result in a srt file
+  -owts,     --output-words      [false  ] output script for generating karaoke video
+  -ocsv,     --output-csv        [false  ] output result in a CSV file
+  -of FNAME, --output-file FNAME [       ] output file path (without file extension)
+  -ps,       --print-special     [false  ] print special tokens
+  -pc,       --print-colors      [false  ] print colors
+  -pp,       --print-progress    [false  ] print progress
+  -nt,       --no-timestamps     [true   ] do not print timestamps
+  -l LANG,   --language LANG     [en     ] spoken language ('auto' for auto-detect)
+             --prompt PROMPT     [       ] initial prompt
+  -m FNAME,  --model FNAME       [models/ggml-base.en.bin] model path
+  -f FNAME,  --file FNAME        [       ] input WAV file path
 
 
 bash ./models/download-ggml-model.sh base.en
@@ -137,7 +141,8 @@ Running base.en on all samples in ./samples ...
 [+] Running base.en on samples/jfk.wav ... (run 'ffplay samples/jfk.wav' to listen)
 ----------------------------------------------
 
-whisper_model_load: loading model from 'models/ggml-base.en.bin'
+whisper_init_from_file: loading model from 'models/ggml-base.en.bin'
+whisper_model_load: loading model
 whisper_model_load: n_vocab       = 51864
 whisper_model_load: n_audio_ctx   = 1500
 whisper_model_load: n_audio_state = 512
@@ -150,13 +155,14 @@ whisper_model_load: n_text_layer  = 6
 whisper_model_load: n_mels        = 80
 whisper_model_load: f16           = 1
 whisper_model_load: type          = 2
+whisper_model_load: mem required  =  215.00 MB (+    6.00 MB per decoder)
+whisper_model_load: kv self size  =    5.25 MB
+whisper_model_load: kv cross size =   17.58 MB
 whisper_model_load: adding 1607 extra tokens
-whisper_model_load: mem_required  =  506.00 MB
-whisper_model_load: ggml ctx size =  140.60 MB
-whisper_model_load: memory size   =   22.83 MB
+whisper_model_load: model ctx     =  140.60 MB
 whisper_model_load: model size    =  140.54 MB
 
-system_info: n_threads = 4 / 10 | AVX = 0 | AVX2 = 0 | AVX512 = 0 | NEON = 1 | FP16_VA = 1 | WASM_SIMD = 0 | BLAS = 1 |
+system_info: n_threads = 4 / 10 | AVX = 0 | AVX2 = 0 | AVX512 = 0 | FMA = 0 | NEON = 1 | ARM_FMA = 1 | F16C = 0 | FP16_VA = 1 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 0 | VSX = 0 |
 
 main: processing 'samples/jfk.wav' (176000 samples, 11.0 sec), 4 threads, 1 processors, lang = en, task = transcribe, timestamps = 1 ...
 
@@ -164,12 +170,13 @@ main: processing 'samples/jfk.wav' (176000 samples, 11.0 sec), 4 threads, 1 proc
 [00:00:00.000 --> 00:00:11.000]   And so my fellow Americans, ask not what your country can do for you, ask what you can do for your country.
 
 
-whisper_print_timings:     load time =   105.91 ms
-whisper_print_timings:      mel time =    24.62 ms
-whisper_print_timings:   sample time =     3.63 ms
-whisper_print_timings:   encode time =   324.71 ms / 54.12 ms per layer
-whisper_print_timings:   decode time =    83.58 ms / 13.93 ms per layer
-whisper_print_timings:    total time =   542.81 ms
+whisper_print_timings:     fallbacks =   0 p /   0 h
+whisper_print_timings:     load time =   113.81 ms
+whisper_print_timings:      mel time =    15.40 ms
+whisper_print_timings:   sample time =    11.58 ms /    27 runs (    0.43 ms per run)
+whisper_print_timings:   encode time =   266.60 ms /     1 runs (  266.60 ms per run)
+whisper_print_timings:   decode time =    66.11 ms /    27 runs (    2.45 ms per run)
+whisper_print_timings:    total time =   476.31 ms
 ```
 
 The command downloads the `base.en` model converted to custom `ggml` format and runs the inference on all `.wav` samples in the folder `samples`.
@@ -212,12 +219,66 @@ make large
 
 | Model  | Disk   | Mem     | SHA                                        |
 | ---    | ---    | ---     | ---                                        |
-| tiny   |  75 MB | ~390 MB | `bd577a113a864445d4c299885e0cb97d4ba92b5f` |
-| base   | 142 MB | ~500 MB | `465707469ff3a37a2b9b8d8f89f2f99de7299dac` |
-| small  | 466 MB | ~1.0 GB | `55356645c2b361a969dfd0ef2c5a50d530afd8d5` |
-| medium | 1.5 GB | ~2.6 GB | `fd9727b6e1217c2f614f9b698455c4ffd82463b4` |
-| large  | 2.9 GB | ~4.7 GB | `0f4c8e34f21cf1a914c59d8b3ce882345ad349d6` |
+| tiny   |  75 MB | ~125 MB | `bd577a113a864445d4c299885e0cb97d4ba92b5f` |
+| base   | 142 MB | ~210 MB | `465707469ff3a37a2b9b8d8f89f2f99de7299dac` |
+| small  | 466 MB | ~600 MB | `55356645c2b361a969dfd0ef2c5a50d530afd8d5` |
+| medium | 1.5 GB | ~1.7 GB | `fd9727b6e1217c2f614f9b698455c4ffd82463b4` |
+| large  | 2.9 GB | ~3.3 GB | `0f4c8e34f21cf1a914c59d8b3ce882345ad349d6` |
 
+## Core ML support
+
+On Apple Silicon devices, the Encoder inference can be executed on the Apple Neural Engine (ANE) via Core ML. This can result in significant
+speed-up - more than x3 faster compared with CPU-only execution. Here are the instructions for generating a Core ML model and using it with `whisper.cpp`:
+
+- Install Python dependencies needed for the creation of the Core ML model:
+
+  ```bash
+  pip install ane_transformers
+  pip install openai-whisper
+  pip install coremltools
+  ```
+
+- Generate a Core ML model. For example, to generate a `base.en` model, use:
+
+  ```bash
+  ./models/generate-coreml-model.sh base.en
+  ```
+
+  This will generate the folder `models/ggml-base.en-encoder.mlmodelc`
+
+- Build `whisper.cpp` with Core ML support:
+
+  ```bash
+  # using Makefile
+  make clean
+  WHISPER_COREML=1 make -j
+  
+  # using CMake
+  cd build
+  cmake -DWHISPER_COREML=1 ..
+  ```
+
+- Run the examples as usual. For example:
+
+  ```bash
+  ./main -m models/ggml-base.en.bin -f samples/jfk.wav
+
+  ...
+
+  whisper_init_state: loading Core ML model from 'models/ggml-base.en-encoder.mlmodelc'
+  whisper_init_state: first run on a device may take a while ...
+  whisper_init_state: Core ML model loaded
+
+  system_info: n_threads = 4 / 10 | AVX = 0 | AVX2 = 0 | AVX512 = 0 | FMA = 0 | NEON = 1 | ARM_FMA = 1 | F16C = 0 | FP16_VA = 1 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 0 | VSX = 0 | COREML = 1 | 
+
+  ...
+  ```
+
+  The first run on a device is slow, since the ANE service compiles the Core ML model to some device-specific format.
+  Next runs are faster.
+  
+For more information about the Core ML implementation please refer to PR [#566](https://github.com/ggerganov/whisper.cpp/pull/566).
+  
 ## Limitations
 
 - Inference only
@@ -234,7 +295,8 @@ in about half a minute on a MacBook M1 Pro, using `medium.en` model:
 ```java
 $ ./main -m models/ggml-medium.en.bin -f samples/gb1.wav -t 8
 
-whisper_model_load: loading model from 'models/ggml-medium.en.bin'
+whisper_init_from_file: loading model from 'models/ggml-medium.en.bin'
+whisper_model_load: loading model
 whisper_model_load: n_vocab       = 51864
 whisper_model_load: n_audio_ctx   = 1500
 whisper_model_load: n_audio_state = 1024
@@ -247,62 +309,67 @@ whisper_model_load: n_text_layer  = 24
 whisper_model_load: n_mels        = 80
 whisper_model_load: f16           = 1
 whisper_model_load: type          = 4
-whisper_model_load: mem_required  = 2610.00 MB
+whisper_model_load: mem required  = 1720.00 MB (+   43.00 MB per decoder)
+whisper_model_load: kv self size  =   42.00 MB
+whisper_model_load: kv cross size =  140.62 MB
 whisper_model_load: adding 1607 extra tokens
-whisper_model_load: ggml ctx size = 1644.97 MB
-whisper_model_load: memory size =   182.62 MB
-whisper_model_load: model size  =  1462.12 MB
+whisper_model_load: model ctx     = 1462.35 MB
+whisper_model_load: model size    = 1462.12 MB
 
-main: processing 'samples/gb1.wav' (3179750 samples, 198.7 sec), 8 threads, lang = en, task = transcribe, timestamps = 1 ...
+system_info: n_threads = 8 / 10 | AVX = 0 | AVX2 = 0 | AVX512 = 0 | FMA = 0 | NEON = 1 | ARM_FMA = 1 | F16C = 0 | FP16_VA = 1 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 0 | VSX = 0 |
 
-[00:00.000 --> 00:08.000]   My fellow Americans, this day has brought terrible news and great sadness to our country.
-[00:08.000 --> 00:17.000]   At nine o'clock this morning, Mission Control in Houston lost contact with our Space Shuttle Columbia.
-[00:17.000 --> 00:23.000]   A short time later, debris was seen falling from the skies above Texas.
-[00:23.000 --> 00:29.000]   The Columbia's lost. There are no survivors.
-[00:29.000 --> 00:32.000]   On board was a crew of seven.
-[00:32.000 --> 00:39.000]   Colonel Rick Husband, Lieutenant Colonel Michael Anderson, Commander Laurel Clark,
-[00:39.000 --> 00:48.000]   Captain David Brown, Commander William McCool, Dr. Kultna Shavla, and Ilan Ramon,
-[00:48.000 --> 00:52.000]   a colonel in the Israeli Air Force.
-[00:52.000 --> 00:58.000]   These men and women assumed great risk in the service to all humanity.
-[00:58.000 --> 01:03.000]   In an age when space flight has come to seem almost routine,
-[01:03.000 --> 01:07.000]   it is easy to overlook the dangers of travel by rocket
-[01:07.000 --> 01:12.000]   and the difficulties of navigating the fierce outer atmosphere of the Earth.
-[01:12.000 --> 01:18.000]   These astronauts knew the dangers, and they faced them willingly,
-[01:18.000 --> 01:23.000]   knowing they had a high and noble purpose in life.
-[01:23.000 --> 01:31.000]   Because of their courage and daring and idealism, we will miss them all the more.
-[01:31.000 --> 01:36.000]   All Americans today are thinking as well of the families of these men and women
-[01:36.000 --> 01:40.000]   who have been given this sudden shock and grief.
-[01:40.000 --> 01:45.000]   You're not alone. Our entire nation grieves with you,
-[01:45.000 --> 01:52.000]   and those you love will always have the respect and gratitude of this country.
-[01:52.000 --> 01:56.000]   The cause in which they died will continue.
-[01:56.000 --> 02:04.000]   Mankind is led into the darkness beyond our world by the inspiration of discovery
-[02:04.000 --> 02:11.000]   and the longing to understand. Our journey into space will go on.
-[02:11.000 --> 02:16.000]   In the skies today, we saw destruction and tragedy.
-[02:16.000 --> 02:22.000]   Yet farther than we can see, there is comfort and hope.
-[02:22.000 --> 02:29.000]   In the words of the prophet Isaiah, "Lift your eyes and look to the heavens
-[02:29.000 --> 02:35.000]   who created all these. He who brings out the starry hosts one by one
-[02:35.000 --> 02:39.000]   and calls them each by name."
-[02:39.000 --> 02:46.000]   Because of His great power and mighty strength, not one of them is missing.
-[02:46.000 --> 02:55.000]   The same Creator who names the stars also knows the names of the seven souls we mourn today.
-[02:55.000 --> 03:01.000]   The crew of the shuttle Columbia did not return safely to earth,
-[03:01.000 --> 03:05.000]   yet we can pray that all are safely home.
-[03:05.000 --> 03:13.000]   May God bless the grieving families, and may God continue to bless America.
-[03:13.000 --> 03:41.000]   Audio
+main: processing 'samples/gb1.wav' (3179750 samples, 198.7 sec), 8 threads, 1 processors, lang = en, task = transcribe, timestamps = 1 ...
 
 
-whisper_print_timings:     load time =   575.92 ms
-whisper_print_timings:      mel time =   230.60 ms
-whisper_print_timings:   sample time =    73.19 ms
-whisper_print_timings:   encode time = 19552.61 ms / 814.69 ms per layer
-whisper_print_timings:   decode time = 13249.96 ms / 552.08 ms per layer
-whisper_print_timings:    total time = 33686.27 ms
+[00:00:00.000 --> 00:00:08.000]   My fellow Americans, this day has brought terrible news and great sadness to our country.
+[00:00:08.000 --> 00:00:17.000]   At nine o'clock this morning, Mission Control in Houston lost contact with our Space Shuttle Columbia.
+[00:00:17.000 --> 00:00:23.000]   A short time later, debris was seen falling from the skies above Texas.
+[00:00:23.000 --> 00:00:29.000]   The Columbia's lost. There are no survivors.
+[00:00:29.000 --> 00:00:32.000]   On board was a crew of seven.
+[00:00:32.000 --> 00:00:39.000]   Colonel Rick Husband, Lieutenant Colonel Michael Anderson, Commander Laurel Clark,
+[00:00:39.000 --> 00:00:48.000]   Captain David Brown, Commander William McCool, Dr. Kultna Shavla, and Ilan Ramon,
+[00:00:48.000 --> 00:00:52.000]   a colonel in the Israeli Air Force.
+[00:00:52.000 --> 00:00:58.000]   These men and women assumed great risk in the service to all humanity.
+[00:00:58.000 --> 00:01:03.000]   In an age when space flight has come to seem almost routine,
+[00:01:03.000 --> 00:01:07.000]   it is easy to overlook the dangers of travel by rocket
+[00:01:07.000 --> 00:01:12.000]   and the difficulties of navigating the fierce outer atmosphere of the Earth.
+[00:01:12.000 --> 00:01:18.000]   These astronauts knew the dangers, and they faced them willingly,
+[00:01:18.000 --> 00:01:23.000]   knowing they had a high and noble purpose in life.
+[00:01:23.000 --> 00:01:31.000]   Because of their courage and daring and idealism, we will miss them all the more.
+[00:01:31.000 --> 00:01:36.000]   All Americans today are thinking as well of the families of these men and women
+[00:01:36.000 --> 00:01:40.000]   who have been given this sudden shock and grief.
+[00:01:40.000 --> 00:01:45.000]   You're not alone. Our entire nation grieves with you,
+[00:01:45.000 --> 00:01:52.000]   and those you love will always have the respect and gratitude of this country.
+[00:01:52.000 --> 00:01:56.000]   The cause in which they died will continue.
+[00:01:56.000 --> 00:02:04.000]   Mankind is led into the darkness beyond our world by the inspiration of discovery
+[00:02:04.000 --> 00:02:11.000]   and the longing to understand. Our journey into space will go on.
+[00:02:11.000 --> 00:02:16.000]   In the skies today, we saw destruction and tragedy.
+[00:02:16.000 --> 00:02:22.000]   Yet farther than we can see, there is comfort and hope.
+[00:02:22.000 --> 00:02:29.000]   In the words of the prophet Isaiah, "Lift your eyes and look to the heavens
+[00:02:29.000 --> 00:02:35.000]   who created all these. He who brings out the starry hosts one by one
+[00:02:35.000 --> 00:02:39.000]   and calls them each by name."
+[00:02:39.000 --> 00:02:46.000]   Because of His great power and mighty strength, not one of them is missing.
+[00:02:46.000 --> 00:02:55.000]   The same Creator who names the stars also knows the names of the seven souls we mourn today.
+[00:02:55.000 --> 00:03:01.000]   The crew of the shuttle Columbia did not return safely to earth,
+[00:03:01.000 --> 00:03:05.000]   yet we can pray that all are safely home.
+[00:03:05.000 --> 00:03:13.000]   May God bless the grieving families, and may God continue to bless America.
+[00:03:13.000 --> 00:03:19.000]   [Silence]
+
+
+whisper_print_timings:     fallbacks =   1 p /   0 h
+whisper_print_timings:     load time =   569.03 ms
+whisper_print_timings:      mel time =   146.85 ms
+whisper_print_timings:   sample time =   238.66 ms /   553 runs (    0.43 ms per run)
+whisper_print_timings:   encode time = 18665.10 ms /     9 runs ( 2073.90 ms per run)
+whisper_print_timings:   decode time = 13090.93 ms /   549 runs (   23.85 ms per run)
+whisper_print_timings:    total time = 32733.52 ms
 ```
 </details>
 
 ## Real-time audio input example
 
 This is a naive example of performing real-time inference on audio from your microphone.
-The [stream](examples/stream) tool samples the audio every half a second and runs the transcription continously.
+The [stream](examples/stream) tool samples the audio every half a second and runs the transcription continuously.
 More info is available in [issue #10](https://github.com/ggerganov/whisper.cpp/issues/10).
 
 ```java
@@ -317,18 +384,22 @@ https://user-images.githubusercontent.com/1991296/194935793-76afede7-cfa8-48d8-a
 Adding the `--print-colors` argument will print the transcribed text using an experimental color coding strategy
 to highlight words with high or low confidence:
 
+```java
+./main -m models/ggml-base.en.bin -f samples/gb0.wav --print-colors
+```
+
 <img width="965" alt="image" src="https://user-images.githubusercontent.com/1991296/197356445-311c8643-9397-4e5e-b46e-0b4b4daa2530.png">
 
 ## Controlling the length of the generated text segments (experimental)
 
-For example, to limit the line length to a maximum of 16 characters, simply add `-ml 16`: 
+For example, to limit the line length to a maximum of 16 characters, simply add `-ml 16`:
 
 ```java
 ./main -m ./models/ggml-base.en.bin -f ./samples/jfk.wav -ml 16
 
 whisper_model_load: loading model from './models/ggml-base.en.bin'
 ...
-system_info: n_threads = 4 / 10 | AVX2 = 0 | AVX512 = 0 | NEON = 1 | FP16_VA = 1 | WASM_SIMD = 0 | BLAS = 1 | 
+system_info: n_threads = 4 / 10 | AVX2 = 0 | AVX512 = 0 | NEON = 1 | FP16_VA = 1 | WASM_SIMD = 0 | BLAS = 1 |
 
 main: processing './samples/jfk.wav' (176000 samples, 11.0 sec), 4 threads, 1 processors, lang = en, task = transcribe, timestamps = 1 ...
 
@@ -352,7 +423,7 @@ The `--max-len` argument can be used to obtain word-level timestamps. Simply use
 
 whisper_model_load: loading model from './models/ggml-base.en.bin'
 ...
-system_info: n_threads = 4 / 10 | AVX2 = 0 | AVX512 = 0 | NEON = 1 | FP16_VA = 1 | WASM_SIMD = 0 | BLAS = 1 | 
+system_info: n_threads = 4 / 10 | AVX2 = 0 | AVX512 = 0 | NEON = 1 | FP16_VA = 1 | WASM_SIMD = 0 | BLAS = 1 |
 
 main: processing './samples/jfk.wav' (176000 samples, 11.0 sec), 4 threads, 1 processors, lang = en, task = transcribe, timestamps = 1 ...
 
@@ -422,6 +493,19 @@ https://user-images.githubusercontent.com/1991296/199337538-b7b0c7a3-2753-4a88-a
 
 ---
 
+## Video comparison of different models
+
+Use the [extra/bench-wts.sh](https://github.com/ggerganov/whisper.cpp/blob/master/extra/bench-wts.sh) script to generate a video in the following format:
+
+```java
+./extra/bench-wts.sh samples/jfk.wav
+ffplay ./samples/jfk.wav.all.mp4
+```
+
+https://user-images.githubusercontent.com/1991296/223206245-2d36d903-cf8e-4f09-8c3b-eb9f9c39d6fc.mp4
+
+---
+
 ## Benchmarks
 
 In order to have an objective comparison of the performance of the inference across different system configurations,
@@ -442,7 +526,7 @@ The original models are converted to a custom binary format. This allows to pack
 You can download the converted models using the [models/download-ggml-model.sh](models/download-ggml-model.sh) script
 or manually from here:
 
-- https://huggingface.co/datasets/ggerganov/whisper.cpp
+- https://huggingface.co/ggerganov/whisper.cpp
 - https://ggml.ggerganov.com
 
 For more details, see the conversion script [models/convert-pt-to-ggml.py](models/convert-pt-to-ggml.py) or the README
@@ -452,9 +536,19 @@ in [models](models).
 
 - [X] Rust: [tazz4843/whisper-rs](https://github.com/tazz4843/whisper-rs) | [#310](https://github.com/ggerganov/whisper.cpp/discussions/310)
 - [X] Javascript: [bindings/javascript](bindings/javascript) | [#309](https://github.com/ggerganov/whisper.cpp/discussions/309)
+  - React Native (iOS / Android): [whisper.rn](https://github.com/mybigday/whisper.rn)
 - [X] Go: [bindings/go](bindings/go) | [#312](https://github.com/ggerganov/whisper.cpp/discussions/312)
+- [X] Ruby: [bindings/ruby](bindings/ruby) | [#507](https://github.com/ggerganov/whisper.cpp/discussions/507)
 - [X] Objective-C / Swift: [ggerganov/whisper.spm](https://github.com/ggerganov/whisper.spm) | [#313](https://github.com/ggerganov/whisper.cpp/discussions/313)
-- [ ] Python: soon | [WIP](https://github.com/ggerganov/whisper.cpp/issues/9)
+  - [exPHAT/SwiftWhisper](https://github.com/exPHAT/SwiftWhisper)
+- [X] .NET: | [#422](https://github.com/ggerganov/whisper.cpp/discussions/422)
+  - [sandrohanea/whisper.net](https://github.com/sandrohanea/whisper.net)
+  - [NickDarvey/whisper](https://github.com/NickDarvey/whisper)
+- [X] Python: | [#9](https://github.com/ggerganov/whisper.cpp/issues/9)
+  - [stlukey/whispercpp.py](https://github.com/stlukey/whispercpp.py) (Cython)
+  - [aarnphm/whispercpp](https://github.com/aarnphm/whispercpp) (Pybind11)
+- [X] R: [bnosac/audio.whisper](https://github.com/bnosac/audio.whisper)
+- [X] Unity: [macoron/whisper.unity](https://github.com/Macoron/whisper.unity)
 
 ## Examples
 
@@ -468,6 +562,7 @@ Some of the examples are even ported to run in the browser using WebAssembly. Ch
 | [stream](examples/stream) | [stream.wasm](examples/stream.wasm) | Real-time transcription of raw microphone capture |
 | [command](examples/command) | [command.wasm](examples/command.wasm) | Basic voice assistant example for receiving voice commands from the mic |
 | [talk](examples/talk) | [talk.wasm](examples/talk.wasm) | Talk with a GPT-2 bot |
+| [talk-llama](examples/talk-llama) | | Talk with a LLaMA bot |
 | [whisper.objc](examples/whisper.objc) | | iOS mobile application using whisper.cpp |
 | [whisper.swiftui](examples/whisper.swiftui) | | SwiftUI iOS / macOS application using whisper.cpp |
 | [whisper.android](examples/whisper.android) | | Android mobile application using whisper.cpp |

bindings/go/examples/go-model-download/main.go

@@ -17,9 +17,9 @@ import (
 // CONSTANTS
 
 const (
-	srcUrl  = "https://huggingface.co/datasets/ggerganov/whisper.cpp/resolve/main" // The location of the models
-	srcExt  = ".bin"                                                               // Filename extension
-	bufSize = 1024 * 64                                                            // Size of the buffer used for downloading the model
+	srcUrl  = "https://huggingface.co/ggerganov/whisper.cpp/resolve/main" // The location of the models
+	srcExt  = ".bin"                                                      // Filename extension
+	bufSize = 1024 * 64                                                   // Size of the buffer used for downloading the model
 )
 
 var (

bindings/go/examples/go-whisper/process.go

@@ -25,6 +25,8 @@ func Process(model whisper.Model, path string, flags *Flags) error {
 		return err
 	}
 
+	fmt.Printf("\n%s\n", context.SystemInfo())
+
 	// Open the file
 	fmt.Fprintf(flags.Output(), "Loading %q\n", path)
 	fh, err := os.Open(path)
@@ -64,10 +66,13 @@ func Process(model whisper.Model, path string, flags *Flags) error {
 
 	// Process the data
 	fmt.Fprintf(flags.Output(), "  ...processing %q\n", path)
+	context.ResetTimings()
 	if err := context.Process(data, cb); err != nil {
 		return err
 	}
 
+	context.PrintTimings()
+
 	// Print out the results
 	switch {
 	case flags.GetOut() == "srt":

bindings/go/params.go

@@ -49,6 +49,10 @@ func (p *Params) SetSpeedup(v bool) {
 
 // Set language id
 func (p *Params) SetLanguage(lang int) error {
+	if lang == -1 {
+		p.language = nil
+		return nil
+	}
 	str := C.whisper_lang_str(C.int(lang))
 	if str == nil {
 		return ErrInvalidLanguage
@@ -66,6 +70,11 @@ func (p *Params) Language() int {
 	return int(C.whisper_lang_id(p.language))
 }
 
+// Threads available
+func (p *Params) Threads() int {
+	return int(p.n_threads)
+}
+
 // Set number of threads to use
 func (p *Params) SetThreads(threads int) {
 	p.n_threads = C.int(threads)
@@ -96,6 +105,10 @@ func (p *Params) SetMaxSegmentLength(n int) {
 	p.max_len = C.int(n)
 }
 
+func (p *Params) SetTokenTimestamps(b bool) {
+	p.token_timestamps = toBool(b)
+}
+
 // Set max tokens per segment (0 = no limit)
 func (p *Params) SetMaxTokensPerSegment(n int) {
 	p.max_tokens = C.int(n)

bindings/go/pkg/whisper/context.go

@@ -1,7 +1,9 @@
 package whisper
 
 import (
+	"fmt"
 	"io"
+	"runtime"
 	"strings"
 	"time"
 
@@ -44,7 +46,10 @@ func (context *context) SetLanguage(lang string) error {
 	if !context.model.IsMultilingual() {
 		return ErrModelNotMultilingual
 	}
-	if id := context.model.ctx.Whisper_lang_id(lang); id < 0 {
+
+	if lang == "auto" {
+		context.params.SetLanguage(-1)
+	} else if id := context.model.ctx.Whisper_lang_id(lang); id < 0 {
 		return ErrUnsupportedLanguage
 	} else if err := context.params.SetLanguage(id); err != nil {
 		return err
@@ -59,6 +64,10 @@ func (context *context) IsMultilingual() bool {
 
 // Get language
 func (context *context) Language() string {
+	id := context.params.Language()
+	if id == -1 {
+		return "auto"
+	}
 	return whisper.Whisper_lang_str(context.params.Language())
 }
 
@@ -102,11 +111,46 @@ func (context *context) SetMaxSegmentLength(n uint) {
 	context.params.SetMaxSegmentLength(int(n))
 }
 
+// Set token timestamps flag
+func (context *context) SetTokenTimestamps(b bool) {
+	context.params.SetTokenTimestamps(b)
+}
+
 // Set max tokens per segment (0 = no limit)
 func (context *context) SetMaxTokensPerSegment(n uint) {
 	context.params.SetMaxTokensPerSegment(int(n))
 }
 
+// ResetTimings resets the mode timings. Should be called before processing
+func (context *context) ResetTimings() {
+	context.model.ctx.Whisper_reset_timings()
+}
+
+// PrintTimings prints the model timings to stdout.
+func (context *context) PrintTimings() {
+	context.model.ctx.Whisper_print_timings()
+}
+
+// SystemInfo returns the system information
+func (context *context) SystemInfo() string {
+	return fmt.Sprintf("system_info: n_threads = %d / %d | %s\n",
+		context.params.Threads(),
+		runtime.NumCPU(),
+		whisper.Whisper_print_system_info(),
+	)
+}
+
+// Use mel data at offset_ms to try and auto-detect the spoken language
+// Make sure to call whisper_pcm_to_mel() or whisper_set_mel() first.
+// Returns the probabilities of all languages.
+func (context *context) WhisperLangAutoDetect(offset_ms int, n_threads int) ([]float32, error) {
+	langProbs, err := context.model.ctx.Whisper_lang_auto_detect(offset_ms, n_threads)
+	if err != nil {
+		return nil, err
+	}
+	return langProbs, nil
+}
+
 // Process new sample data and return any errors
 func (context *context) Process(data []float32, cb SegmentCallback) error {
 	if context.model.ctx == nil {
@@ -241,10 +285,14 @@ func toSegment(ctx *whisper.Context, n int) Segment {
 func toTokens(ctx *whisper.Context, n int) []Token {
 	result := make([]Token, ctx.Whisper_full_n_tokens(n))
 	for i := 0; i < len(result); i++ {
+		data := ctx.Whisper_full_get_token_data(n, i)
+
 		result[i] = Token{
-			Id:   int(ctx.Whisper_full_get_token_id(n, i)),
-			Text: strings.TrimSpace(ctx.Whisper_full_get_token_text(n, i)),
-			P:    ctx.Whisper_full_get_token_p(n, i),
+			Id:    int(ctx.Whisper_full_get_token_id(n, i)),
+			Text:  ctx.Whisper_full_get_token_text(n, i),
+			P:     ctx.Whisper_full_get_token_p(n, i),
+			Start: time.Duration(data.T0()) * time.Millisecond * 10,
+			End:   time.Duration(data.T1()) * time.Millisecond * 10,
 		}
 	}
 	return result

bindings/go/pkg/whisper/interface.go

@@ -29,7 +29,7 @@ type Model interface {
 
 // Context is the speach recognition context.
 type Context interface {
-	SetLanguage(string) error // Set the language to use for speech recognition.
+	SetLanguage(string) error // Set the language to use for speech recognition, use "auto" for auto detect language.
 	SetTranslate(bool)        // Set translate flag
 	IsMultilingual() bool     // Return true if the model is multilingual.
 	Language() string         // Get language
@@ -41,6 +41,7 @@ type Context interface {
 	SetTokenThreshold(float32)    // Set timestamp token probability threshold
 	SetTokenSumThreshold(float32) // Set timestamp token sum probability threshold
 	SetMaxSegmentLength(uint)     // Set max segment length in characters
+	SetTokenTimestamps(bool)      // Set token timestamps flag
 	SetMaxTokensPerSegment(uint)  // Set max tokens per segment (0 = no limit)
 
 	// Process mono audio data and return any errors.
@@ -60,6 +61,12 @@ type Context interface {
 	IsNOT(Token) bool          // Test for "No timestamps" token
 	IsLANG(Token, string) bool // Test for token associated with a specific language
 	IsText(Token) bool         // Test for text token
+
+	// Timings
+	PrintTimings()
+	ResetTimings()
+
+	SystemInfo() string
 }
 
 // Segment is the text result of a speech recognition.
@@ -79,7 +86,8 @@ type Segment struct {
 
 // Token is a text or special token
 type Token struct {
-	Id   int
-	Text string
-	P    float32
+	Id         int
+	Text       string
+	P          float32
+	Start, End time.Duration
 }

bindings/go/pkg/whisper/model.go

@@ -94,6 +94,7 @@ func (model *model) NewContext() (Context, error) {
 	params.SetPrintRealtime(false)
 	params.SetPrintTimestamps(false)
 	params.SetThreads(runtime.NumCPU())
+	params.SetNoContext(true)
 
 	// Return new context
 	return newContext(model, params)

bindings/go/whisper.go

@@ -20,7 +20,7 @@ extern bool callEncoderBegin(void* user_data);
 // Text segment callback
 // Called on every newly generated text segment
 // Use the whisper_full_...() functions to obtain the text segments
-static void whisper_new_segment_cb(struct whisper_context* ctx, int n_new, void* user_data) {
+static void whisper_new_segment_cb(struct whisper_context* ctx, struct whisper_state* state, int n_new, void* user_data) {
     if(user_data != NULL && ctx != NULL) {
         callNewSegment(user_data, n_new);
     }
@@ -29,7 +29,7 @@ static void whisper_new_segment_cb(struct whisper_context* ctx, int n_new, void*
 // Encoder begin callback
 // If not NULL, called before the encoder starts
 // If it returns false, the computation is aborted
-static bool whisper_encoder_begin_cb(struct whisper_context* ctx, void* user_data) {
+static bool whisper_encoder_begin_cb(struct whisper_context* ctx, struct whisper_state* state, void* user_data) {
     if(user_data != NULL && ctx != NULL) {
         return callEncoderBegin(user_data);
     }
@@ -147,16 +147,6 @@ func (ctx *Context) Whisper_decode(tokens []Token, past, threads int) error {
 	}
 }
 
-// whisper_sample_best() returns the token with the highest probability
-func (ctx *Context) Whisper_sample_best() TokenData {
-	return TokenData(C.whisper_sample_best((*C.struct_whisper_context)(ctx)))
-}
-
-// whisper_sample_timestamp() returns the most probable timestamp token
-func (ctx *Context) Whisper_sample_timestamp(is_initial bool) TokenData {
-	return TokenData(C.whisper_sample_timestamp((*C.struct_whisper_context)(ctx), C.bool(is_initial)))
-}
-
 // Convert the provided text into tokens. The tokens pointer must be large enough to hold the resulting tokens.
 // Returns the number of tokens on success
 func (ctx *Context) Whisper_tokenize(text string, tokens []Token) (int, error) {
@@ -366,7 +356,7 @@ func (ctx *Context) Whisper_full_get_token_id(segment int, token int) Token {
 
 // Get token data for the specified token in the specified segment.
 // This contains probabilities, timestamps, etc.
-func (ctx *Context) whisper_full_get_token_data(segment int, token int) TokenData {
+func (ctx *Context) Whisper_full_get_token_data(segment int, token int) TokenData {
 	return TokenData(C.whisper_full_get_token_data((*C.struct_whisper_context)(ctx), C.int(segment), C.int(token)))
 }
 
@@ -417,3 +407,11 @@ func callEncoderBegin(user_data unsafe.Pointer) C.bool {
 	}
 	return true
 }
+
+func (t TokenData) T0() int64 {
+	return int64(t.t0)
+}
+
+func (t TokenData) T1() int64 {
+	return int64(t.t1)
+}

bindings/javascript/package.json

@@ -1,6 +1,6 @@
 {
   "name": "whisper.cpp",
-  "version": "1.1.0",
+  "version": "1.3.0",
   "description": "Whisper speech recognition",
   "main": "whisper.js",
   "scripts": {

bindings/ruby/ext/.gitignore

@@ -0,0 +1,7 @@
+Makefile
+ggml.c
+ggml.h
+whisper.bundle
+whisper.cpp
+whisper.h
+dr_wav.h

bindings/ruby/ext/extconf.rb

@@ -0,0 +1,21 @@
+require 'mkmf'
+system("cp #{File.join(File.dirname(__FILE__),'..','..','..','whisper.cpp')} .")
+system("cp #{File.join(File.dirname(__FILE__),'..','..','..','whisper.h')} .")
+system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml.h')} .")
+system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml.c')} .")
+system("cp #{File.join(File.dirname(__FILE__),'..','..','..','examples','dr_wav.h')} .")
+
+
+# need to use c++ compiler flags
+$CXXFLAGS << ' -std=c++11'
+# Set to true when building binary gems
+if enable_config('static-stdlib', false)
+  $LDFLAGS << ' -static-libgcc -static-libstdc++'
+end
+
+if enable_config('march-tune-native', false)
+  $CFLAGS << ' -march=native -mtune=native'
+  $CXXFLAGS << ' -march=native -mtune=native'
+end
+
+create_makefile('whisper')

bindings/ruby/ext/ruby_whisper.cpp

@@ -0,0 +1,426 @@
+#include <ruby.h>
+#include "ruby_whisper.h"
+#define DR_WAV_IMPLEMENTATION
+#include "dr_wav.h"
+#include <cmath>
+#include <fstream>
+#include <cstdio>
+#include <string>
+#include <thread>
+#include <vector>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define BOOL_PARAMS_SETTER(self, prop, value) \
+  ruby_whisper_params *rwp; \
+  Data_Get_Struct(self, ruby_whisper_params, rwp); \
+  if (value == Qfalse || value == Qnil) { \
+    rwp->params.prop = false; \
+  } else { \
+    rwp->params.prop = true; \
+  } \
+  return value; \
+
+#define BOOL_PARAMS_GETTER(self,  prop) \
+  ruby_whisper_params *rwp; \
+  Data_Get_Struct(self, ruby_whisper_params, rwp); \
+  if (rwp->params.prop) { \
+    return Qtrue; \
+  } else { \
+    return Qfalse; \
+  }
+
+VALUE mWhisper;
+VALUE cContext;
+VALUE cParams;
+
+static void ruby_whisper_free(ruby_whisper *rw) {
+  if (rw->context) {
+    whisper_free(rw->context);
+    rw->context = NULL;
+  }
+}
+static void ruby_whisper_params_free(ruby_whisper_params *rwp) {
+}
+
+void rb_whisper_mark(ruby_whisper *rw) {
+  // call rb_gc_mark on any ruby references in rw
+}
+
+void rb_whisper_free(ruby_whisper *rw) {
+  ruby_whisper_free(rw);
+  free(rw);
+}
+
+void rb_whisper_params_mark(ruby_whisper_params *rwp) {
+}
+
+void rb_whisper_params_free(ruby_whisper_params *rwp) {
+  ruby_whisper_params_free(rwp);
+  free(rwp);
+}
+
+static VALUE ruby_whisper_allocate(VALUE klass) {
+  ruby_whisper *rw;
+  rw = ALLOC(ruby_whisper);
+  rw->context = NULL;
+  return Data_Wrap_Struct(klass, rb_whisper_mark, rb_whisper_free, rw);
+}
+
+static VALUE ruby_whisper_params_allocate(VALUE klass) {
+  ruby_whisper_params *rwp;
+  rwp = ALLOC(ruby_whisper_params);
+  rwp->params = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
+  return Data_Wrap_Struct(klass, rb_whisper_params_mark, rb_whisper_params_free, rwp);
+}
+
+static VALUE ruby_whisper_initialize(int argc, VALUE *argv, VALUE self) {
+  ruby_whisper *rw;
+  VALUE whisper_model_file_path;
+
+  // TODO: we can support init from buffer here too maybe another ruby object to expose
+  rb_scan_args(argc, argv, "01", &whisper_model_file_path);
+  Data_Get_Struct(self, ruby_whisper, rw);
+
+  if (!rb_respond_to(whisper_model_file_path, rb_intern("to_s"))) {
+    rb_raise(rb_eRuntimeError, "Expected file path to model to initialize Whisper::Context");
+  }
+  rw->context = whisper_init_from_file(StringValueCStr(whisper_model_file_path));
+  if (rw->context == nullptr) {
+    rb_raise(rb_eRuntimeError, "error: failed to initialize whisper context");
+  }
+  return self;
+}
+
+/*
+ * transcribe a single file
+ * can emit to a block results
+ *
+ **/
+static VALUE ruby_whisper_transcribe(int argc, VALUE *argv, VALUE self) {
+  ruby_whisper *rw;
+  ruby_whisper_params *rwp;
+  VALUE wave_file_path, blk, params;
+
+  rb_scan_args(argc, argv, "02&", &wave_file_path, &params, &blk);
+  Data_Get_Struct(self, ruby_whisper, rw);
+  Data_Get_Struct(params, ruby_whisper_params, rwp);
+
+  if (!rb_respond_to(wave_file_path, rb_intern("to_s"))) {
+    rb_raise(rb_eRuntimeError, "Expected file path to wave file");
+  }
+
+  std::string fname_inp = StringValueCStr(wave_file_path);
+
+  std::vector<float> pcmf32; // mono-channel F32 PCM
+  std::vector<std::vector<float>> pcmf32s; // stereo-channel F32 PCM
+
+  // WAV input - this is directly from main.cpp example
+  {
+    drwav wav;
+    std::vector<uint8_t> wav_data; // used for pipe input from stdin
+
+    if (fname_inp == "-") {
+      {
+        uint8_t buf[1024];
+        while (true) {
+          const size_t n = fread(buf, 1, sizeof(buf), stdin);
+          if (n == 0) {
+            break;
+          }
+          wav_data.insert(wav_data.end(), buf, buf + n);
+        }
+      }
+
+      if (drwav_init_memory(&wav, wav_data.data(), wav_data.size(), nullptr) == false) {
+        fprintf(stderr, "error: failed to open WAV file from stdin\n");
+        return self;
+      }
+
+      fprintf(stderr, "%s: read %zu bytes from stdin\n", __func__, wav_data.size());
+    } else if (drwav_init_file(&wav, fname_inp.c_str(), nullptr) == false) {
+      fprintf(stderr, "error: failed to open '%s' as WAV file\n", fname_inp.c_str());
+      return self;
+    }
+
+    if (wav.channels != 1 && wav.channels != 2) {
+      fprintf(stderr, "WAV file '%s' must be mono or stereo\n", fname_inp.c_str());
+      return self;
+    }
+
+    if (rwp->diarize && wav.channels != 2 && rwp->params.print_timestamps == false) {
+      fprintf(stderr, "WAV file '%s' must be stereo for diarization and timestamps have to be enabled\n", fname_inp.c_str());
+      return self;
+    }
+
+    if (wav.sampleRate != WHISPER_SAMPLE_RATE) {
+      fprintf(stderr, "WAV file '%s' must be %i kHz\n", fname_inp.c_str(), WHISPER_SAMPLE_RATE/1000);
+      return self;
+    }
+
+    if (wav.bitsPerSample != 16) {
+      fprintf(stderr, "WAV file '%s' must be 16-bit\n", fname_inp.c_str());
+      return self;
+    }
+
+    const uint64_t n = wav_data.empty() ? wav.totalPCMFrameCount : wav_data.size()/(wav.channels*wav.bitsPerSample/8);
+
+    std::vector<int16_t> pcm16;
+    pcm16.resize(n*wav.channels);
+    drwav_read_pcm_frames_s16(&wav, n, pcm16.data());
+    drwav_uninit(&wav);
+
+    // convert to mono, float
+    pcmf32.resize(n);
+    if (wav.channels == 1) {
+      for (uint64_t i = 0; i < n; i++) {
+        pcmf32[i] = float(pcm16[i])/32768.0f;
+      }
+    } else {
+      for (uint64_t i = 0; i < n; i++) {
+        pcmf32[i] = float(pcm16[2*i] + pcm16[2*i + 1])/65536.0f;
+      }
+    }
+
+    if (rwp->diarize) {
+      // convert to stereo, float
+      pcmf32s.resize(2);
+
+      pcmf32s[0].resize(n);
+      pcmf32s[1].resize(n);
+      for (uint64_t i = 0; i < n; i++) {
+        pcmf32s[0][i] = float(pcm16[2*i])/32768.0f;
+        pcmf32s[1][i] = float(pcm16[2*i + 1])/32768.0f;
+      }
+    }
+  }
+  {
+    static bool is_aborted = false; // NOTE: this should be atomic to avoid data race
+
+    rwp->params.encoder_begin_callback = [](struct whisper_context * /*ctx*/, struct whisper_state * /*state*/, void * user_data) {
+      bool is_aborted = *(bool*)user_data;
+      return !is_aborted;
+    };
+    rwp->params.encoder_begin_callback_user_data = &is_aborted;
+  }
+
+  if (whisper_full_parallel(rw->context, rwp->params, pcmf32.data(), pcmf32.size(), 1) != 0) {
+    fprintf(stderr, "failed to process audio\n");
+    return self;
+  }
+  const int n_segments = whisper_full_n_segments(rw->context);
+  VALUE output = rb_str_new2("");
+  for (int i = 0; i < n_segments; ++i) {
+    const char * text = whisper_full_get_segment_text(rw->context, i);
+    output = rb_str_concat(output, rb_str_new2(text));
+  }
+  VALUE idCall = rb_intern("call");
+  if (blk != Qnil) {
+    rb_funcall(blk, idCall, 1, output);
+  }
+  return self;
+}
+
+/*
+ * params.language = "auto" | "en", etc...
+ */
+static VALUE ruby_whisper_params_set_language(VALUE self, VALUE value) {
+  ruby_whisper_params *rwp;
+  Data_Get_Struct(self, ruby_whisper_params, rwp);
+  if (value == Qfalse || value == Qnil) {
+    rwp->params.language = "auto";
+  } else {
+    rwp->params.language = StringValueCStr(value);
+  }
+  return value;
+}
+static VALUE ruby_whisper_params_get_language(VALUE self) {
+  ruby_whisper_params *rwp;
+  Data_Get_Struct(self, ruby_whisper_params, rwp);
+  if (rwp->params.language) {
+    return rb_str_new2(rwp->params.language);
+  } else {
+    return rb_str_new2("auto");
+  }
+}
+static VALUE ruby_whisper_params_set_translate(VALUE self, VALUE value) {
+  BOOL_PARAMS_SETTER(self, translate, value)
+}
+static VALUE ruby_whisper_params_get_translate(VALUE self) {
+  BOOL_PARAMS_GETTER(self, translate)
+}
+static VALUE ruby_whisper_params_set_no_context(VALUE self, VALUE value) {
+  BOOL_PARAMS_SETTER(self, no_context, value)
+}
+static VALUE ruby_whisper_params_get_no_context(VALUE self) {
+  BOOL_PARAMS_GETTER(self, no_context)
+}
+static VALUE ruby_whisper_params_set_single_segment(VALUE self, VALUE value) {
+  BOOL_PARAMS_SETTER(self, single_segment, value)
+}
+static VALUE ruby_whisper_params_get_single_segment(VALUE self) {
+  BOOL_PARAMS_GETTER(self, single_segment)
+}
+static VALUE ruby_whisper_params_set_print_special(VALUE self, VALUE value) {
+  BOOL_PARAMS_SETTER(self, print_special, value)
+}
+static VALUE ruby_whisper_params_get_print_special(VALUE self) {
+  BOOL_PARAMS_GETTER(self, print_special)
+}
+static VALUE ruby_whisper_params_set_print_progress(VALUE self, VALUE value) {
+  BOOL_PARAMS_SETTER(self, print_progress, value)
+}
+static VALUE ruby_whisper_params_get_print_progress(VALUE self) {
+  BOOL_PARAMS_GETTER(self, print_progress)
+}
+static VALUE ruby_whisper_params_set_print_realtime(VALUE self, VALUE value) {
+  BOOL_PARAMS_SETTER(self, print_realtime, value)
+}
+static VALUE ruby_whisper_params_get_print_realtime(VALUE self) {
+  BOOL_PARAMS_GETTER(self, print_realtime)
+}
+static VALUE ruby_whisper_params_set_print_timestamps(VALUE self, VALUE value) {
+  BOOL_PARAMS_SETTER(self, print_timestamps, value)
+}
+static VALUE ruby_whisper_params_get_print_timestamps(VALUE self) {
+  BOOL_PARAMS_GETTER(self, print_timestamps)
+}
+static VALUE ruby_whisper_params_set_suppress_blank(VALUE self, VALUE value) {
+  BOOL_PARAMS_SETTER(self, suppress_blank, value)
+}
+static VALUE ruby_whisper_params_get_suppress_blank(VALUE self) {
+  BOOL_PARAMS_GETTER(self, suppress_blank)
+}
+static VALUE ruby_whisper_params_set_suppress_non_speech_tokens(VALUE self, VALUE value) {
+  BOOL_PARAMS_SETTER(self, suppress_non_speech_tokens, value)
+}
+static VALUE ruby_whisper_params_get_suppress_non_speech_tokens(VALUE self) {
+  BOOL_PARAMS_GETTER(self, suppress_non_speech_tokens)
+}
+static VALUE ruby_whisper_params_get_token_timestamps(VALUE self) {
+  BOOL_PARAMS_GETTER(self, token_timestamps)
+}
+static VALUE ruby_whisper_params_set_token_timestamps(VALUE self, VALUE value) {
+  BOOL_PARAMS_SETTER(self, token_timestamps, value)
+}
+static VALUE ruby_whisper_params_get_split_on_word(VALUE self) {
+  BOOL_PARAMS_GETTER(self, split_on_word)
+}
+static VALUE ruby_whisper_params_set_split_on_word(VALUE self, VALUE value) {
+  BOOL_PARAMS_SETTER(self, split_on_word, value)
+}
+static VALUE ruby_whisper_params_get_speed_up(VALUE self) {
+  BOOL_PARAMS_GETTER(self, speed_up)
+}
+static VALUE ruby_whisper_params_set_speed_up(VALUE self, VALUE value) {
+  BOOL_PARAMS_SETTER(self, speed_up, value)
+}
+static VALUE ruby_whisper_params_get_diarize(VALUE self) {
+  ruby_whisper_params *rwp;
+  Data_Get_Struct(self, ruby_whisper_params, rwp);
+  if (rwp->diarize) {
+    return Qtrue;
+  } else {
+    return Qfalse;
+  }
+}
+static VALUE ruby_whisper_params_set_diarize(VALUE self, VALUE value) {
+  ruby_whisper_params *rwp;
+  Data_Get_Struct(self, ruby_whisper_params, rwp);
+  if (value == Qfalse || value == Qnil) {
+    rwp->diarize = false;
+  } else {
+    rwp->diarize = true;
+  } \
+  return value;
+}
+
+static VALUE ruby_whisper_params_get_offset(VALUE self) {
+  ruby_whisper_params *rwp;
+  Data_Get_Struct(self, ruby_whisper_params, rwp);
+  return INT2NUM(rwp->params.offset_ms);
+}
+static VALUE ruby_whisper_params_set_offset(VALUE self, VALUE value) {
+  ruby_whisper_params *rwp;
+  Data_Get_Struct(self, ruby_whisper_params, rwp);
+  rwp->params.offset_ms = NUM2INT(value);
+  return value;
+}
+static VALUE ruby_whisper_params_get_duration(VALUE self) {
+  ruby_whisper_params *rwp;
+  Data_Get_Struct(self, ruby_whisper_params, rwp);
+  return INT2NUM(rwp->params.duration_ms);
+}
+static VALUE ruby_whisper_params_set_duration(VALUE self, VALUE value) {
+  ruby_whisper_params *rwp;
+  Data_Get_Struct(self, ruby_whisper_params, rwp);
+  rwp->params.duration_ms = NUM2INT(value);
+  return value;
+}
+
+static VALUE ruby_whisper_params_get_max_text_tokens(VALUE self) {
+  ruby_whisper_params *rwp;
+  Data_Get_Struct(self, ruby_whisper_params, rwp);
+  return INT2NUM(rwp->params.n_max_text_ctx);
+}
+static VALUE ruby_whisper_params_set_max_text_tokens(VALUE self, VALUE value) {
+  ruby_whisper_params *rwp;
+  Data_Get_Struct(self, ruby_whisper_params, rwp);
+  rwp->params.n_max_text_ctx = NUM2INT(value);
+  return value;
+}
+
+void Init_whisper() {
+  mWhisper = rb_define_module("Whisper");
+  cContext = rb_define_class_under(mWhisper, "Context", rb_cObject);
+  cParams  = rb_define_class_under(mWhisper, "Params", rb_cObject);
+
+  rb_define_alloc_func(cContext, ruby_whisper_allocate);
+  rb_define_method(cContext, "initialize", ruby_whisper_initialize, -1);
+
+  rb_define_method(cContext, "transcribe", ruby_whisper_transcribe, -1);
+
+  rb_define_alloc_func(cParams, ruby_whisper_params_allocate);
+
+  rb_define_method(cParams, "language=", ruby_whisper_params_set_language, 1);
+  rb_define_method(cParams, "language", ruby_whisper_params_get_language, 0);
+  rb_define_method(cParams, "translate=", ruby_whisper_params_set_translate, 1);
+  rb_define_method(cParams, "translate", ruby_whisper_params_get_translate, 0);
+  rb_define_method(cParams, "no_context=", ruby_whisper_params_set_no_context, 1);
+  rb_define_method(cParams, "no_context", ruby_whisper_params_get_no_context, 0);
+  rb_define_method(cParams, "single_segment=", ruby_whisper_params_set_single_segment, 1);
+  rb_define_method(cParams, "single_segment", ruby_whisper_params_get_single_segment, 0);
+  rb_define_method(cParams, "print_special", ruby_whisper_params_get_print_special, 0);
+  rb_define_method(cParams, "print_special=", ruby_whisper_params_set_print_special, 1);
+  rb_define_method(cParams, "print_progress", ruby_whisper_params_get_print_progress, 0);
+  rb_define_method(cParams, "print_progress=", ruby_whisper_params_set_print_progress, 1);
+  rb_define_method(cParams, "print_realtime", ruby_whisper_params_get_print_realtime, 0);
+  rb_define_method(cParams, "print_realtime=", ruby_whisper_params_set_print_realtime, 1);
+  rb_define_method(cParams, "print_timestamps", ruby_whisper_params_get_print_timestamps, 0);
+  rb_define_method(cParams, "print_timestamps=", ruby_whisper_params_set_print_timestamps, 1);
+  rb_define_method(cParams, "suppress_blank", ruby_whisper_params_get_suppress_blank, 0);
+  rb_define_method(cParams, "suppress_blank=", ruby_whisper_params_set_suppress_blank, 1);
+  rb_define_method(cParams, "suppress_non_speech_tokens", ruby_whisper_params_get_suppress_non_speech_tokens, 0);
+  rb_define_method(cParams, "suppress_non_speech_tokens=", ruby_whisper_params_set_suppress_non_speech_tokens, 1);
+  rb_define_method(cParams, "token_timestamps", ruby_whisper_params_get_token_timestamps, 0);
+  rb_define_method(cParams, "token_timestamps=", ruby_whisper_params_set_token_timestamps, 1);
+  rb_define_method(cParams, "split_on_word", ruby_whisper_params_get_split_on_word, 0);
+  rb_define_method(cParams, "split_on_word=", ruby_whisper_params_set_split_on_word, 1);
+  rb_define_method(cParams, "speed_up", ruby_whisper_params_get_speed_up, 0);
+  rb_define_method(cParams, "speed_up=", ruby_whisper_params_set_speed_up, 1);
+  rb_define_method(cParams, "diarize", ruby_whisper_params_get_diarize, 0);
+  rb_define_method(cParams, "diarize=", ruby_whisper_params_set_diarize, 1);
+
+  rb_define_method(cParams, "offset", ruby_whisper_params_get_offset, 0);
+  rb_define_method(cParams, "offset=", ruby_whisper_params_set_offset, 1);
+  rb_define_method(cParams, "duration", ruby_whisper_params_get_duration, 0);
+  rb_define_method(cParams, "duration=", ruby_whisper_params_set_duration, 1);
+
+  rb_define_method(cParams, "max_text_tokens", ruby_whisper_params_get_max_text_tokens, 0);
+  rb_define_method(cParams, "max_text_tokens=", ruby_whisper_params_set_max_text_tokens, 1);
+}
+#ifdef __cplusplus
+}
+#endif

bindings/ruby/ext/ruby_whisper.h

@@ -0,0 +1,15 @@
+#ifndef __RUBY_WHISPER_H
+#define __RUBY_WHISPER_H
+
+#include "whisper.h"
+
+typedef struct {
+  struct whisper_context *context;
+} ruby_whisper;
+
+typedef struct {
+  struct whisper_full_params params;
+  bool diarize;
+} ruby_whisper_params;
+
+#endif

bindings/ruby/tests/test_whisper.rb

@@ -0,0 +1,138 @@
+TOPDIR = File.expand_path(File.join(File.dirname(__FILE__), '..'))
+EXTDIR = File.join(TOPDIR, 'ext')
+#$LIBDIR = File.join(TOPDIR, 'lib')
+#$:.unshift(LIBDIR)
+$:.unshift(EXTDIR)
+
+require 'whisper'
+require 'test/unit'
+
+class TestWhisper < Test::Unit::TestCase
+  def setup
+    @params  = Whisper::Params.new
+  end
+
+  def test_language
+    @params.language = "en"
+    assert_equal @params.language, "en"
+    @params.language = "auto"
+    assert_equal @params.language, "auto"
+  end
+
+  def test_offset
+    @params.offset = 10_000
+    assert_equal @params.offset, 10_000
+    @params.offset = 0
+    assert_equal @params.offset, 0
+  end
+
+  def test_duration
+    @params.duration = 60_000
+    assert_equal @params.duration, 60_000
+    @params.duration = 0
+    assert_equal @params.duration, 0
+  end
+
+  def test_max_text_tokens
+    @params.max_text_tokens = 300
+    assert_equal @params.max_text_tokens, 300
+    @params.max_text_tokens = 0
+    assert_equal @params.max_text_tokens, 0
+  end
+
+  def test_translate
+    @params.translate = true
+    assert @params.translate
+    @params.translate = false
+    assert !@params.translate
+  end
+
+  def test_no_context
+    @params.no_context = true
+    assert @params.no_context
+    @params.no_context = false
+    assert !@params.no_context
+  end
+
+  def test_single_segment
+    @params.single_segment = true
+    assert @params.single_segment
+    @params.single_segment = false
+    assert !@params.single_segment
+  end
+
+  def test_print_special
+    @params.print_special = true
+    assert @params.print_special
+    @params.print_special = false
+    assert !@params.print_special
+  end
+
+  def test_print_progress
+    @params.print_progress = true
+    assert @params.print_progress
+    @params.print_progress = false
+    assert !@params.print_progress
+  end
+
+  def test_print_realtime
+    @params.print_realtime = true
+    assert @params.print_realtime
+    @params.print_realtime = false
+    assert !@params.print_realtime
+  end
+
+  def test_print_timestamps
+    @params.print_timestamps = true
+    assert @params.print_timestamps
+    @params.print_timestamps = false
+    assert !@params.print_timestamps
+  end
+
+  def test_suppress_blank
+    @params.suppress_blank = true
+    assert @params.suppress_blank
+    @params.suppress_blank = false
+    assert !@params.suppress_blank
+  end
+
+  def test_suppress_non_speech_tokens
+    @params.suppress_non_speech_tokens = true
+    assert @params.suppress_non_speech_tokens
+    @params.suppress_non_speech_tokens = false
+    assert !@params.suppress_non_speech_tokens
+  end
+
+  def test_token_timestamps
+    @params.token_timestamps = true
+    assert @params.token_timestamps
+    @params.token_timestamps = false
+    assert !@params.token_timestamps
+  end
+
+  def test_split_on_word
+    @params.split_on_word = true
+    assert @params.split_on_word
+    @params.split_on_word = false
+    assert !@params.split_on_word
+  end
+
+  def test_speed_up
+    @params.speed_up = true
+    assert @params.speed_up
+    @params.speed_up = false
+    assert !@params.speed_up
+  end
+
+  def test_whisper
+    @whisper = Whisper::Context.new(File.join(TOPDIR, '..', '..', 'models', 'ggml-base.en.bin'))
+    params  = Whisper::Params.new
+    params.print_timestamps = false
+
+    jfk = File.join(TOPDIR, '..', '..', 'samples', 'jfk.wav')
+    @whisper.transcribe(jfk, params) {|text|
+      assert_match /ask not what your country can do for you, ask what you can do for your country/, text
+    }
+  end
+
+end

coreml/whisper-decoder-impl.h

@@ -0,0 +1,146 @@
+//
+// whisper-decoder-impl.h
+//
+// This file was automatically generated and should not be edited.
+//
+
+#import <Foundation/Foundation.h>
+#import <CoreML/CoreML.h>
+#include <stdint.h>
+#include <os/log.h>
+
+NS_ASSUME_NONNULL_BEGIN
+
+
+/// Model Prediction Input Type
+API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((visibility("hidden")))
+@interface whisper_decoder_implInput : NSObject<MLFeatureProvider>
+
+/// token_data as 1 by 1 matrix of 32-bit integers
+@property (readwrite, nonatomic, strong) MLMultiArray * token_data;
+
+/// audio_data as 1 × 384 × 1 × 1500 4-dimensional array of floats
+@property (readwrite, nonatomic, strong) MLMultiArray * audio_data;
+- (instancetype)init NS_UNAVAILABLE;
+- (instancetype)initWithToken_data:(MLMultiArray *)token_data audio_data:(MLMultiArray *)audio_data NS_DESIGNATED_INITIALIZER;
+
+@end
+
+
+/// Model Prediction Output Type
+API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((visibility("hidden")))
+@interface whisper_decoder_implOutput : NSObject<MLFeatureProvider>
+
+/// var_1346 as multidimensional array of floats
+@property (readwrite, nonatomic, strong) MLMultiArray * var_1346;
+- (instancetype)init NS_UNAVAILABLE;
+- (instancetype)initWithVar_1346:(MLMultiArray *)var_1346 NS_DESIGNATED_INITIALIZER;
+
+@end
+
+
+/// Class for model loading and prediction
+API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((visibility("hidden")))
+@interface whisper_decoder_impl : NSObject
+@property (readonly, nonatomic, nullable) MLModel * model;
+
+/**
+    URL of the underlying .mlmodelc directory.
+*/
++ (nullable NSURL *)URLOfModelInThisBundle;
+
+/**
+    Initialize whisper_decoder_impl instance from an existing MLModel object.
+
+    Usually the application does not use this initializer unless it makes a subclass of whisper_decoder_impl.
+    Such application may want to use `-[MLModel initWithContentsOfURL:configuration:error:]` and `+URLOfModelInThisBundle` to create a MLModel object to pass-in.
+*/
+- (instancetype)initWithMLModel:(MLModel *)model NS_DESIGNATED_INITIALIZER;
+
+/**
+    Initialize whisper_decoder_impl instance with the model in this bundle.
+*/
+- (nullable instancetype)init;
+
+/**
+    Initialize whisper_decoder_impl instance with the model in this bundle.
+
+    @param configuration The model configuration object
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+*/
+- (nullable instancetype)initWithConfiguration:(MLModelConfiguration *)configuration error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Initialize whisper_decoder_impl instance from the model URL.
+
+    @param modelURL URL to the .mlmodelc directory for whisper_decoder_impl.
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+*/
+- (nullable instancetype)initWithContentsOfURL:(NSURL *)modelURL error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Initialize whisper_decoder_impl instance from the model URL.
+
+    @param modelURL URL to the .mlmodelc directory for whisper_decoder_impl.
+    @param configuration The model configuration object
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+*/
+- (nullable instancetype)initWithContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Construct whisper_decoder_impl instance asynchronously with configuration.
+    Model loading may take time when the model content is not immediately available (e.g. encrypted model). Use this factory method especially when the caller is on the main thread.
+
+    @param configuration The model configuration
+    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid whisper_decoder_impl instance or NSError object.
+*/
++ (void)loadWithConfiguration:(MLModelConfiguration *)configuration completionHandler:(void (^)(whisper_decoder_impl * _Nullable model, NSError * _Nullable error))handler;
+
+/**
+    Construct whisper_decoder_impl instance asynchronously with URL of .mlmodelc directory and optional configuration.
+
+    Model loading may take time when the model content is not immediately available (e.g. encrypted model). Use this factory method especially when the caller is on the main thread.
+
+    @param modelURL The model URL.
+    @param configuration The model configuration
+    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid whisper_decoder_impl instance or NSError object.
+*/
++ (void)loadContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration completionHandler:(void (^)(whisper_decoder_impl * _Nullable model, NSError * _Nullable error))handler;
+
+/**
+    Make a prediction using the standard interface
+    @param input an instance of whisper_decoder_implInput to predict from
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+    @return the prediction as whisper_decoder_implOutput
+*/
+- (nullable whisper_decoder_implOutput *)predictionFromFeatures:(whisper_decoder_implInput *)input error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Make a prediction using the standard interface
+    @param input an instance of whisper_decoder_implInput to predict from
+    @param options prediction options
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+    @return the prediction as whisper_decoder_implOutput
+*/
+- (nullable whisper_decoder_implOutput *)predictionFromFeatures:(whisper_decoder_implInput *)input options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Make a prediction using the convenience interface
+    @param token_data as 1 by 1 matrix of 32-bit integers:
+    @param audio_data as 1 × 384 × 1 × 1500 4-dimensional array of floats:
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+    @return the prediction as whisper_decoder_implOutput
+*/
+- (nullable whisper_decoder_implOutput *)predictionFromToken_data:(MLMultiArray *)token_data audio_data:(MLMultiArray *)audio_data error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Batch prediction
+    @param inputArray array of whisper_decoder_implInput instances to obtain predictions from
+    @param options prediction options
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+    @return the predictions as NSArray<whisper_decoder_implOutput *>
+*/
+- (nullable NSArray<whisper_decoder_implOutput *> *)predictionsFromInputs:(NSArray<whisper_decoder_implInput*> *)inputArray options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+@end
+
+NS_ASSUME_NONNULL_END

coreml/whisper-decoder-impl.m

@@ -0,0 +1,201 @@
+//
+// whisper-decoder-impl.m
+//
+// This file was automatically generated and should not be edited.
+//
+
+#if !__has_feature(objc_arc)
+#error This file must be compiled with automatic reference counting enabled (-fobjc-arc)
+#endif
+
+#import "whisper-decoder-impl.h"
+
+@implementation whisper_decoder_implInput
+
+- (instancetype)initWithToken_data:(MLMultiArray *)token_data audio_data:(MLMultiArray *)audio_data {
+    self = [super init];
+    if (self) {
+        _token_data = token_data;
+        _audio_data = audio_data;
+    }
+    return self;
+}
+
+- (NSSet<NSString *> *)featureNames {
+    return [NSSet setWithArray:@[@"token_data", @"audio_data"]];
+}
+
+- (nullable MLFeatureValue *)featureValueForName:(NSString *)featureName {
+    if ([featureName isEqualToString:@"token_data"]) {
+        return [MLFeatureValue featureValueWithMultiArray:self.token_data];
+    }
+    if ([featureName isEqualToString:@"audio_data"]) {
+        return [MLFeatureValue featureValueWithMultiArray:self.audio_data];
+    }
+    return nil;
+}
+
+@end
+
+@implementation whisper_decoder_implOutput
+
+- (instancetype)initWithVar_1346:(MLMultiArray *)var_1346 {
+    self = [super init];
+    if (self) {
+        _var_1346 = var_1346;
+    }
+    return self;
+}
+
+- (NSSet<NSString *> *)featureNames {
+    return [NSSet setWithArray:@[@"var_1346"]];
+}
+
+- (nullable MLFeatureValue *)featureValueForName:(NSString *)featureName {
+    if ([featureName isEqualToString:@"var_1346"]) {
+        return [MLFeatureValue featureValueWithMultiArray:self.var_1346];
+    }
+    return nil;
+}
+
+@end
+
+@implementation whisper_decoder_impl
+
+
+/**
+    URL of the underlying .mlmodelc directory.
+*/
++ (nullable NSURL *)URLOfModelInThisBundle {
+    NSString *assetPath = [[NSBundle bundleForClass:[self class]] pathForResource:@"whisper_decoder_impl" ofType:@"mlmodelc"];
+    if (nil == assetPath) { os_log_error(OS_LOG_DEFAULT, "Could not load whisper-decoder-impl.mlmodelc in the bundle resource"); return nil; }
+    return [NSURL fileURLWithPath:assetPath];
+}
+
+
+/**
+    Initialize whisper_decoder_impl instance from an existing MLModel object.
+
+    Usually the application does not use this initializer unless it makes a subclass of whisper_decoder_impl.
+    Such application may want to use `-[MLModel initWithContentsOfURL:configuration:error:]` and `+URLOfModelInThisBundle` to create a MLModel object to pass-in.
+*/
+- (instancetype)initWithMLModel:(MLModel *)model {
+    self = [super init];
+    if (!self) { return nil; }
+    _model = model;
+    if (_model == nil) { return nil; }
+    return self;
+}
+
+
+/**
+    Initialize whisper_decoder_impl instance with the model in this bundle.
+*/
+- (nullable instancetype)init {
+    return [self initWithContentsOfURL:(NSURL * _Nonnull)self.class.URLOfModelInThisBundle error:nil];
+}
+
+
+/**
+    Initialize whisper_decoder_impl instance with the model in this bundle.
+
+    @param configuration The model configuration object
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+*/
+- (nullable instancetype)initWithConfiguration:(MLModelConfiguration *)configuration error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    return [self initWithContentsOfURL:(NSURL * _Nonnull)self.class.URLOfModelInThisBundle configuration:configuration error:error];
+}
+
+
+/**
+    Initialize whisper_decoder_impl instance from the model URL.
+
+    @param modelURL URL to the .mlmodelc directory for whisper_decoder_impl.
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+*/
+- (nullable instancetype)initWithContentsOfURL:(NSURL *)modelURL error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    MLModel *model = [MLModel modelWithContentsOfURL:modelURL error:error];
+    if (model == nil) { return nil; }
+    return [self initWithMLModel:model];
+}
+
+
+/**
+    Initialize whisper_decoder_impl instance from the model URL.
+
+    @param modelURL URL to the .mlmodelc directory for whisper_decoder_impl.
+    @param configuration The model configuration object
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+*/
+- (nullable instancetype)initWithContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    MLModel *model = [MLModel modelWithContentsOfURL:modelURL configuration:configuration error:error];
+    if (model == nil) { return nil; }
+    return [self initWithMLModel:model];
+}
+
+
+/**
+    Construct whisper_decoder_impl instance asynchronously with configuration.
+    Model loading may take time when the model content is not immediately available (e.g. encrypted model). Use this factory method especially when the caller is on the main thread.
+
+    @param configuration The model configuration
+    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid whisper_decoder_impl instance or NSError object.
+*/
++ (void)loadWithConfiguration:(MLModelConfiguration *)configuration completionHandler:(void (^)(whisper_decoder_impl * _Nullable model, NSError * _Nullable error))handler {
+    [self loadContentsOfURL:(NSURL * _Nonnull)[self URLOfModelInThisBundle]
+              configuration:configuration
+          completionHandler:handler];
+}
+
+
+/**
+    Construct whisper_decoder_impl instance asynchronously with URL of .mlmodelc directory and optional configuration.
+
+    Model loading may take time when the model content is not immediately available (e.g. encrypted model). Use this factory method especially when the caller is on the main thread.
+
+    @param modelURL The model URL.
+    @param configuration The model configuration
+    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid whisper_decoder_impl instance or NSError object.
+*/
++ (void)loadContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration completionHandler:(void (^)(whisper_decoder_impl * _Nullable model, NSError * _Nullable error))handler {
+    [MLModel loadContentsOfURL:modelURL
+                 configuration:configuration
+             completionHandler:^(MLModel *model, NSError *error) {
+        if (model != nil) {
+            whisper_decoder_impl *typedModel = [[whisper_decoder_impl alloc] initWithMLModel:model];
+            handler(typedModel, nil);
+        } else {
+            handler(nil, error);
+        }
+    }];
+}
+
+- (nullable whisper_decoder_implOutput *)predictionFromFeatures:(whisper_decoder_implInput *)input error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    return [self predictionFromFeatures:input options:[[MLPredictionOptions alloc] init] error:error];
+}
+
+- (nullable whisper_decoder_implOutput *)predictionFromFeatures:(whisper_decoder_implInput *)input options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    id<MLFeatureProvider> outFeatures = [self.model predictionFromFeatures:input options:options error:error];
+    if (!outFeatures) { return nil; }
+    return [[whisper_decoder_implOutput alloc] initWithVar_1346:(MLMultiArray *)[outFeatures featureValueForName:@"var_1346"].multiArrayValue];
+}
+
+- (nullable whisper_decoder_implOutput *)predictionFromToken_data:(MLMultiArray *)token_data audio_data:(MLMultiArray *)audio_data error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    whisper_decoder_implInput *input_ = [[whisper_decoder_implInput alloc] initWithToken_data:token_data audio_data:audio_data];
+    return [self predictionFromFeatures:input_ error:error];
+}
+
+- (nullable NSArray<whisper_decoder_implOutput *> *)predictionsFromInputs:(NSArray<whisper_decoder_implInput*> *)inputArray options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    id<MLBatchProvider> inBatch = [[MLArrayBatchProvider alloc] initWithFeatureProviderArray:inputArray];
+    id<MLBatchProvider> outBatch = [self.model predictionsFromBatch:inBatch options:options error:error];
+    if (!outBatch) { return nil; }
+    NSMutableArray<whisper_decoder_implOutput*> *results = [NSMutableArray arrayWithCapacity:(NSUInteger)outBatch.count];
+    for (NSInteger i = 0; i < outBatch.count; i++) {
+        id<MLFeatureProvider> resultProvider = [outBatch featuresAtIndex:i];
+        whisper_decoder_implOutput * result = [[whisper_decoder_implOutput alloc] initWithVar_1346:(MLMultiArray *)[resultProvider featureValueForName:@"var_1346"].multiArrayValue];
+        [results addObject:result];
+    }
+    return results;
+}
+
+@end

coreml/whisper-encoder-impl.h

@@ -0,0 +1,142 @@
+//
+// whisper-encoder-impl.h
+//
+// This file was automatically generated and should not be edited.
+//
+
+#import <Foundation/Foundation.h>
+#import <CoreML/CoreML.h>
+#include <stdint.h>
+#include <os/log.h>
+
+NS_ASSUME_NONNULL_BEGIN
+
+
+/// Model Prediction Input Type
+API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((visibility("hidden")))
+@interface whisper_encoder_implInput : NSObject<MLFeatureProvider>
+
+/// logmel_data as 1 × 80 × 3000 3-dimensional array of floats
+@property (readwrite, nonatomic, strong) MLMultiArray * logmel_data;
+- (instancetype)init NS_UNAVAILABLE;
+- (instancetype)initWithLogmel_data:(MLMultiArray *)logmel_data NS_DESIGNATED_INITIALIZER;
+
+@end
+
+
+/// Model Prediction Output Type
+API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((visibility("hidden")))
+@interface whisper_encoder_implOutput : NSObject<MLFeatureProvider>
+
+/// output as multidimensional array of floats
+@property (readwrite, nonatomic, strong) MLMultiArray * output;
+- (instancetype)init NS_UNAVAILABLE;
+- (instancetype)initWithOutput:(MLMultiArray *)output NS_DESIGNATED_INITIALIZER;
+
+@end
+
+
+/// Class for model loading and prediction
+API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((visibility("hidden")))
+@interface whisper_encoder_impl : NSObject
+@property (readonly, nonatomic, nullable) MLModel * model;
+
+/**
+    URL of the underlying .mlmodelc directory.
+*/
++ (nullable NSURL *)URLOfModelInThisBundle;
+
+/**
+    Initialize whisper_encoder_impl instance from an existing MLModel object.
+
+    Usually the application does not use this initializer unless it makes a subclass of whisper_encoder_impl.
+    Such application may want to use `-[MLModel initWithContentsOfURL:configuration:error:]` and `+URLOfModelInThisBundle` to create a MLModel object to pass-in.
+*/
+- (instancetype)initWithMLModel:(MLModel *)model NS_DESIGNATED_INITIALIZER;
+
+/**
+    Initialize whisper_encoder_impl instance with the model in this bundle.
+*/
+- (nullable instancetype)init;
+
+/**
+    Initialize whisper_encoder_impl instance with the model in this bundle.
+
+    @param configuration The model configuration object
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+*/
+- (nullable instancetype)initWithConfiguration:(MLModelConfiguration *)configuration error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Initialize whisper_encoder_impl instance from the model URL.
+
+    @param modelURL URL to the .mlmodelc directory for whisper_encoder_impl.
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+*/
+- (nullable instancetype)initWithContentsOfURL:(NSURL *)modelURL error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Initialize whisper_encoder_impl instance from the model URL.
+
+    @param modelURL URL to the .mlmodelc directory for whisper_encoder_impl.
+    @param configuration The model configuration object
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+*/
+- (nullable instancetype)initWithContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Construct whisper_encoder_impl instance asynchronously with configuration.
+    Model loading may take time when the model content is not immediately available (e.g. encrypted model). Use this factory method especially when the caller is on the main thread.
+
+    @param configuration The model configuration
+    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid whisper_encoder_impl instance or NSError object.
+*/
++ (void)loadWithConfiguration:(MLModelConfiguration *)configuration completionHandler:(void (^)(whisper_encoder_impl * _Nullable model, NSError * _Nullable error))handler;
+
+/**
+    Construct whisper_encoder_impl instance asynchronously with URL of .mlmodelc directory and optional configuration.
+
+    Model loading may take time when the model content is not immediately available (e.g. encrypted model). Use this factory method especially when the caller is on the main thread.
+
+    @param modelURL The model URL.
+    @param configuration The model configuration
+    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid whisper_encoder_impl instance or NSError object.
+*/
++ (void)loadContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration completionHandler:(void (^)(whisper_encoder_impl * _Nullable model, NSError * _Nullable error))handler;
+
+/**
+    Make a prediction using the standard interface
+    @param input an instance of whisper_encoder_implInput to predict from
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+    @return the prediction as whisper_encoder_implOutput
+*/
+- (nullable whisper_encoder_implOutput *)predictionFromFeatures:(whisper_encoder_implInput *)input error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Make a prediction using the standard interface
+    @param input an instance of whisper_encoder_implInput to predict from
+    @param options prediction options
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+    @return the prediction as whisper_encoder_implOutput
+*/
+- (nullable whisper_encoder_implOutput *)predictionFromFeatures:(whisper_encoder_implInput *)input options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Make a prediction using the convenience interface
+    @param logmel_data as 1 × 80 × 3000 3-dimensional array of floats:
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+    @return the prediction as whisper_encoder_implOutput
+*/
+- (nullable whisper_encoder_implOutput *)predictionFromLogmel_data:(MLMultiArray *)logmel_data error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Batch prediction
+    @param inputArray array of whisper_encoder_implInput instances to obtain predictions from
+    @param options prediction options
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+    @return the predictions as NSArray<whisper_encoder_implOutput *>
+*/
+- (nullable NSArray<whisper_encoder_implOutput *> *)predictionsFromInputs:(NSArray<whisper_encoder_implInput*> *)inputArray options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+@end
+
+NS_ASSUME_NONNULL_END

coreml/whisper-encoder-impl.m

@@ -0,0 +1,197 @@
+//
+// whisper-encoder-impl.m
+//
+// This file was automatically generated and should not be edited.
+//
+
+#if !__has_feature(objc_arc)
+#error This file must be compiled with automatic reference counting enabled (-fobjc-arc)
+#endif
+
+#import "whisper-encoder-impl.h"
+
+@implementation whisper_encoder_implInput
+
+- (instancetype)initWithLogmel_data:(MLMultiArray *)logmel_data {
+    self = [super init];
+    if (self) {
+        _logmel_data = logmel_data;
+    }
+    return self;
+}
+
+- (NSSet<NSString *> *)featureNames {
+    return [NSSet setWithArray:@[@"logmel_data"]];
+}
+
+- (nullable MLFeatureValue *)featureValueForName:(NSString *)featureName {
+    if ([featureName isEqualToString:@"logmel_data"]) {
+        return [MLFeatureValue featureValueWithMultiArray:self.logmel_data];
+    }
+    return nil;
+}
+
+@end
+
+@implementation whisper_encoder_implOutput
+
+- (instancetype)initWithOutput:(MLMultiArray *)output {
+    self = [super init];
+    if (self) {
+        _output = output;
+    }
+    return self;
+}
+
+- (NSSet<NSString *> *)featureNames {
+    return [NSSet setWithArray:@[@"output"]];
+}
+
+- (nullable MLFeatureValue *)featureValueForName:(NSString *)featureName {
+    if ([featureName isEqualToString:@"output"]) {
+        return [MLFeatureValue featureValueWithMultiArray:self.output];
+    }
+    return nil;
+}
+
+@end
+
+@implementation whisper_encoder_impl
+
+
+/**
+    URL of the underlying .mlmodelc directory.
+*/
++ (nullable NSURL *)URLOfModelInThisBundle {
+    NSString *assetPath = [[NSBundle bundleForClass:[self class]] pathForResource:@"whisper_encoder_impl" ofType:@"mlmodelc"];
+    if (nil == assetPath) { os_log_error(OS_LOG_DEFAULT, "Could not load whisper-encoder-impl.mlmodelc in the bundle resource"); return nil; }
+    return [NSURL fileURLWithPath:assetPath];
+}
+
+
+/**
+    Initialize whisper_encoder_impl instance from an existing MLModel object.
+
+    Usually the application does not use this initializer unless it makes a subclass of whisper_encoder_impl.
+    Such application may want to use `-[MLModel initWithContentsOfURL:configuration:error:]` and `+URLOfModelInThisBundle` to create a MLModel object to pass-in.
+*/
+- (instancetype)initWithMLModel:(MLModel *)model {
+    self = [super init];
+    if (!self) { return nil; }
+    _model = model;
+    if (_model == nil) { return nil; }
+    return self;
+}
+
+
+/**
+    Initialize whisper_encoder_impl instance with the model in this bundle.
+*/
+- (nullable instancetype)init {
+    return [self initWithContentsOfURL:(NSURL * _Nonnull)self.class.URLOfModelInThisBundle error:nil];
+}
+
+
+/**
+    Initialize whisper_encoder_impl instance with the model in this bundle.
+
+    @param configuration The model configuration object
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+*/
+- (nullable instancetype)initWithConfiguration:(MLModelConfiguration *)configuration error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    return [self initWithContentsOfURL:(NSURL * _Nonnull)self.class.URLOfModelInThisBundle configuration:configuration error:error];
+}
+
+
+/**
+    Initialize whisper_encoder_impl instance from the model URL.
+
+    @param modelURL URL to the .mlmodelc directory for whisper_encoder_impl.
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+*/
+- (nullable instancetype)initWithContentsOfURL:(NSURL *)modelURL error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    MLModel *model = [MLModel modelWithContentsOfURL:modelURL error:error];
+    if (model == nil) { return nil; }
+    return [self initWithMLModel:model];
+}
+
+
+/**
+    Initialize whisper_encoder_impl instance from the model URL.
+
+    @param modelURL URL to the .mlmodelc directory for whisper_encoder_impl.
+    @param configuration The model configuration object
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+*/
+- (nullable instancetype)initWithContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    MLModel *model = [MLModel modelWithContentsOfURL:modelURL configuration:configuration error:error];
+    if (model == nil) { return nil; }
+    return [self initWithMLModel:model];
+}
+
+
+/**
+    Construct whisper_encoder_impl instance asynchronously with configuration.
+    Model loading may take time when the model content is not immediately available (e.g. encrypted model). Use this factory method especially when the caller is on the main thread.
+
+    @param configuration The model configuration
+    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid whisper_encoder_impl instance or NSError object.
+*/
++ (void)loadWithConfiguration:(MLModelConfiguration *)configuration completionHandler:(void (^)(whisper_encoder_impl * _Nullable model, NSError * _Nullable error))handler {
+    [self loadContentsOfURL:(NSURL * _Nonnull)[self URLOfModelInThisBundle]
+              configuration:configuration
+          completionHandler:handler];
+}
+
+
+/**
+    Construct whisper_encoder_impl instance asynchronously with URL of .mlmodelc directory and optional configuration.
+
+    Model loading may take time when the model content is not immediately available (e.g. encrypted model). Use this factory method especially when the caller is on the main thread.
+
+    @param modelURL The model URL.
+    @param configuration The model configuration
+    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid whisper_encoder_impl instance or NSError object.
+*/
++ (void)loadContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration completionHandler:(void (^)(whisper_encoder_impl * _Nullable model, NSError * _Nullable error))handler {
+    [MLModel loadContentsOfURL:modelURL
+                 configuration:configuration
+             completionHandler:^(MLModel *model, NSError *error) {
+        if (model != nil) {
+            whisper_encoder_impl *typedModel = [[whisper_encoder_impl alloc] initWithMLModel:model];
+            handler(typedModel, nil);
+        } else {
+            handler(nil, error);
+        }
+    }];
+}
+
+- (nullable whisper_encoder_implOutput *)predictionFromFeatures:(whisper_encoder_implInput *)input error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    return [self predictionFromFeatures:input options:[[MLPredictionOptions alloc] init] error:error];
+}
+
+- (nullable whisper_encoder_implOutput *)predictionFromFeatures:(whisper_encoder_implInput *)input options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    id<MLFeatureProvider> outFeatures = [self.model predictionFromFeatures:input options:options error:error];
+    if (!outFeatures) { return nil; }
+    return [[whisper_encoder_implOutput alloc] initWithOutput:(MLMultiArray *)[outFeatures featureValueForName:@"output"].multiArrayValue];
+}
+
+- (nullable whisper_encoder_implOutput *)predictionFromLogmel_data:(MLMultiArray *)logmel_data error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    whisper_encoder_implInput *input_ = [[whisper_encoder_implInput alloc] initWithLogmel_data:logmel_data];
+    return [self predictionFromFeatures:input_ error:error];
+}
+
+- (nullable NSArray<whisper_encoder_implOutput *> *)predictionsFromInputs:(NSArray<whisper_encoder_implInput*> *)inputArray options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    id<MLBatchProvider> inBatch = [[MLArrayBatchProvider alloc] initWithFeatureProviderArray:inputArray];
+    id<MLBatchProvider> outBatch = [self.model predictionsFromBatch:inBatch options:options error:error];
+    if (!outBatch) { return nil; }
+    NSMutableArray<whisper_encoder_implOutput*> *results = [NSMutableArray arrayWithCapacity:(NSUInteger)outBatch.count];
+    for (NSInteger i = 0; i < outBatch.count; i++) {
+        id<MLFeatureProvider> resultProvider = [outBatch featuresAtIndex:i];
+        whisper_encoder_implOutput * result = [[whisper_encoder_implOutput alloc] initWithOutput:(MLMultiArray *)[resultProvider featureValueForName:@"output"].multiArrayValue];
+        [results addObject:result];
+    }
+    return results;
+}
+
+@end

coreml/whisper-encoder.h

@@ -0,0 +1,22 @@
+// Wrapper of the Core ML Whisper Encoder model
+//
+// Code is derived from the work of Github user @wangchou
+// ref: https://github.com/wangchou/callCoreMLFromCpp
+
+#if __cplusplus
+extern "C" {
+#endif
+
+struct whisper_coreml_context;
+
+struct whisper_coreml_context * whisper_coreml_init(const char * path_model);
+void whisper_coreml_free(struct whisper_coreml_context * ctx);
+
+void whisper_coreml_encode(
+        const whisper_coreml_context * ctx,
+                               float * mel,
+                               float * out);
+
+#if __cplusplus
+}
+#endif

coreml/whisper-encoder.mm

@@ -0,0 +1,67 @@
+#import "coreml/whisper-encoder.h"
+#import "coreml/whisper-encoder-impl.h"
+
+#import <CoreML/CoreML.h>
+
+#include <stdlib.h>
+
+#if __cplusplus
+extern "C" {
+#endif
+
+struct whisper_coreml_context {
+    const void * data;
+};
+
+struct whisper_coreml_context * whisper_coreml_init(const char * path_model) {
+    NSString * path_model_str = [[NSString alloc] initWithUTF8String:path_model];
+
+    NSURL * url_model = [NSURL fileURLWithPath: path_model_str];
+
+    const void * data = CFBridgingRetain([[whisper_encoder_impl alloc] initWithContentsOfURL:url_model error:nil]);
+
+    if (data == NULL) {
+        return NULL;
+    }
+
+    whisper_coreml_context * ctx = new whisper_coreml_context;
+
+    ctx->data = data;
+
+    return ctx;
+}
+
+void whisper_coreml_free(struct whisper_coreml_context * ctx) {
+    CFRelease(ctx->data);
+    delete ctx;
+}
+
+void whisper_coreml_encode(
+        const whisper_coreml_context * ctx,
+                               float * mel,
+                               float * out) {
+    MLMultiArray * inMultiArray = [
+        [MLMultiArray alloc] initWithDataPointer: mel
+                                           shape: @[@1, @80, @3000]
+                                        dataType: MLMultiArrayDataTypeFloat32
+                                         strides: @[@(240000), @(3000), @1]
+                                     deallocator: nil
+                                           error: nil
+    ];
+
+    whisper_encoder_implOutput * outCoreML = [(__bridge id) ctx->data predictionFromLogmel_data:inMultiArray error:nil];
+
+    MLMultiArray * outMA = outCoreML.output;
+
+    //NSArray<NSNumber *> * shape = outMA.shape;
+    //NSArray<NSNumber *> * strides = outMA.strides;
+
+    //printf("shape:   %ld %ld %ld %ld\n", [shape[0] longValue], [shape[1] longValue], [shape[2] longValue], [shape[3] longValue]);
+    //printf("strides: %ld %ld %ld %ld\n", [strides[0] longValue], [strides[1] longValue], [strides[2] longValue], [strides[3] longValue]);
+
+    memcpy(out, outMA.dataPointer, outMA.count * sizeof(float));
+}
+
+#if __cplusplus
+}
+#endif

examples/CMakeLists.txt

@@ -14,6 +14,37 @@ if (WHISPER_SUPPORT_SDL2)
     message(STATUS "SDL2_LIBRARIES = ${SDL2_LIBRARIES}")
 endif()
 
+# common
+
+set(TARGET common)
+
+add_library(${TARGET} STATIC
+    common.h
+    common.cpp
+    )
+
+include(DefaultTargetOptions)
+
+set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+
+if (WHISPER_SUPPORT_SDL2)
+    # common-sdl
+
+    set(TARGET common-sdl)
+
+    add_library(${TARGET} STATIC
+        common-sdl.h
+        common-sdl.cpp
+        )
+
+    include(DefaultTargetOptions)
+
+    target_include_directories(${TARGET} PUBLIC ${SDL2_INCLUDE_DIRS})
+    target_link_libraries(${TARGET} PRIVATE ${SDL2_LIBRARIES})
+
+    set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+endif()
+
 # examples
 
 include_directories(${CMAKE_CURRENT_SOURCE_DIR})
@@ -24,10 +55,13 @@ if (EMSCRIPTEN)
     add_subdirectory(command.wasm)
     add_subdirectory(talk.wasm)
     add_subdirectory(bench.wasm)
+elseif(CMAKE_JS_VERSION)
+    add_subdirectory(addon.node)
 else()
     add_subdirectory(main)
     add_subdirectory(stream)
     add_subdirectory(command)
     add_subdirectory(bench)
     add_subdirectory(talk)
+    add_subdirectory(talk-llama)
 endif()

examples/addon.node/.gitignore

@@ -0,0 +1,3 @@
+.idea
+node_modules
+build

examples/addon.node/CMakeLists.txt

@@ -0,0 +1,31 @@
+set(TARGET whisper-addon)
+
+# Base settings
+#==================================================================
+# env var supported by cmake-js
+add_definitions(-DNAPI_VERSION=4)
+include_directories(${CMAKE_JS_INC})
+#==================================================================
+
+add_library(${TARGET} SHARED ${CMAKE_JS_SRC} addon.cpp)
+set_target_properties(${TARGET} PROPERTIES PREFIX "" SUFFIX ".node")
+
+include(DefaultTargetOptions)
+
+# Include N-API wrappers
+#==================================================================
+execute_process(COMMAND node -p "require('node-addon-api').include"
+        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
+        OUTPUT_VARIABLE NODE_ADDON_API_DIR
+        )
+string(REPLACE "\n" "" NODE_ADDON_API_DIR ${NODE_ADDON_API_DIR})
+string(REPLACE "\"" "" NODE_ADDON_API_DIR ${NODE_ADDON_API_DIR})
+target_include_directories(${TARGET} PRIVATE ${NODE_ADDON_API_DIR})
+#==================================================================
+
+target_link_libraries(${TARGET} ${CMAKE_JS_LIB} common whisper ${CMAKE_THREAD_LIBS_INIT})
+
+if(MSVC AND CMAKE_JS_NODELIB_DEF AND CMAKE_JS_NODELIB_TARGET)
+    # Generate node.lib
+    execute_process(COMMAND ${CMAKE_AR} /def:${CMAKE_JS_NODELIB_DEF} /out:${CMAKE_JS_NODELIB_TARGET} ${CMAKE_STATIC_LINKER_FLAGS})
+endif()

examples/addon.node/README.md

@@ -0,0 +1,37 @@
+# addon
+
+This is an addon demo that can **perform whisper model reasoning in `node` and `electron` environments**, based on [cmake-js](https://github.com/cmake-js/cmake-js).
+It can be used as a reference for using the whisper.cpp project in other node projects.
+
+## Install
+
+```shell
+npm install
+```
+
+## Compile
+
+Make sure it is in the project root directory and compiled with make-js.
+
+```shell
+npx cmake-js compile -T whisper-addon -B Release
+```
+
+For Electron addon and cmake-js options, you can see [cmake-js](https://github.com/cmake-js/cmake-js) and make very few configuration changes.
+
+> Such as appointing special cmake path:
+> ```shell
+> npx cmake-js compile -c 'xxx/cmake' -T whisper-addon -B Release
+> ```
+
+## Run
+
+```shell
+cd examples/addon.node
+
+node index.js --language='language' --model='model-path' --fname_inp='file-path'
+```
+
+Because this is a simple Demo, only the above parameters are set in the node environment.
+
+Other parameters can also be specified in the node environment.

examples/addon.node/__test__/whisper.spec.js

@@ -0,0 +1,22 @@
+const path = require("path");
+const { whisper } = require(path.join(
+  __dirname,
+  "../../../build/Release/whisper-addon"
+));
+const { promisify } = require("util");
+
+const whisperAsync = promisify(whisper);
+
+const whisperParamsMock = {
+  language: "en",
+  model: path.join(__dirname, "../../../models/ggml-base.en.bin"),
+  fname_inp: path.join(__dirname, "../../../samples/jfk.wav"),
+};
+
+describe("Run whisper.node", () => {
+  test("it should receive a non-empty value", async () => {
+    let result = await whisperAsync(whisperParamsMock);
+
+    expect(result.length).toBeGreaterThan(0);
+  });
+});

examples/addon.node/addon.cpp

@@ -0,0 +1,338 @@
+#include "napi.h"
+#include "common.h"
+
+#include "whisper.h"
+
+#include <string>
+#include <thread>
+#include <vector>
+#include <cmath>
+#include <cstdint>
+
+struct whisper_params {
+    int32_t n_threads    = std::min(4, (int32_t) std::thread::hardware_concurrency());
+    int32_t n_processors = 1;
+    int32_t offset_t_ms  = 0;
+    int32_t offset_n     = 0;
+    int32_t duration_ms  = 0;
+    int32_t max_context  = -1;
+    int32_t max_len      = 0;
+    int32_t best_of      = 5;
+    int32_t beam_size    = -1;
+
+    float word_thold    = 0.01f;
+    float entropy_thold = 2.4f;
+    float logprob_thold = -1.0f;
+
+    bool speed_up       = false;
+    bool translate      = false;
+    bool diarize        = false;
+    bool output_txt     = false;
+    bool output_vtt     = false;
+    bool output_srt     = false;
+    bool output_wts     = false;
+    bool output_csv     = false;
+    bool print_special  = false;
+    bool print_colors   = false;
+    bool print_progress = false;
+    bool no_timestamps  = false;
+
+    std::string language = "en";
+    std::string prompt;
+    std::string model    = "../../ggml-large.bin";
+
+    std::vector<std::string> fname_inp = {};
+    std::vector<std::string> fname_out = {};
+};
+
+struct whisper_print_user_data {
+    const whisper_params * params;
+
+    const std::vector<std::vector<float>> * pcmf32s;
+};
+
+//  500 -> 00:05.000
+// 6000 -> 01:00.000
+std::string to_timestamp(int64_t t, bool comma = false) {
+    int64_t msec = t * 10;
+    int64_t hr = msec / (1000 * 60 * 60);
+    msec = msec - hr * (1000 * 60 * 60);
+    int64_t min = msec / (1000 * 60);
+    msec = msec - min * (1000 * 60);
+    int64_t sec = msec / 1000;
+    msec = msec - sec * 1000;
+
+    char buf[32];
+    snprintf(buf, sizeof(buf), "%02d:%02d:%02d%s%03d", (int) hr, (int) min, (int) sec, comma ? "," : ".", (int) msec);
+
+    return std::string(buf);
+}
+
+int timestamp_to_sample(int64_t t, int n_samples) {
+    return std::max(0, std::min((int) n_samples - 1, (int) ((t*WHISPER_SAMPLE_RATE)/100)));
+}
+
+void whisper_print_segment_callback(struct whisper_context * ctx, struct whisper_state * state, int n_new, void * user_data) {
+    const auto & params  = *((whisper_print_user_data *) user_data)->params;
+    const auto & pcmf32s = *((whisper_print_user_data *) user_data)->pcmf32s;
+
+    const int n_segments = whisper_full_n_segments(ctx);
+
+    std::string speaker = "";
+
+    int64_t t0;
+    int64_t t1;
+
+    // print the last n_new segments
+    const int s0 = n_segments - n_new;
+
+    if (s0 == 0) {
+        printf("\n");
+    }
+
+    for (int i = s0; i < n_segments; i++) {
+        if (!params.no_timestamps || params.diarize) {
+            t0 = whisper_full_get_segment_t0(ctx, i);
+            t1 = whisper_full_get_segment_t1(ctx, i);
+        }
+
+        if (!params.no_timestamps) {
+            printf("[%s --> %s]  ", to_timestamp(t0).c_str(), to_timestamp(t1).c_str());
+        }
+
+        if (params.diarize && pcmf32s.size() == 2) {
+            const int64_t n_samples = pcmf32s[0].size();
+
+            const int64_t is0 = timestamp_to_sample(t0, n_samples);
+            const int64_t is1 = timestamp_to_sample(t1, n_samples);
+
+            double energy0 = 0.0f;
+            double energy1 = 0.0f;
+
+            for (int64_t j = is0; j < is1; j++) {
+                energy0 += fabs(pcmf32s[0][j]);
+                energy1 += fabs(pcmf32s[1][j]);
+            }
+
+            if (energy0 > 1.1*energy1) {
+                speaker = "(speaker 0)";
+            } else if (energy1 > 1.1*energy0) {
+                speaker = "(speaker 1)";
+            } else {
+                speaker = "(speaker ?)";
+            }
+
+            //printf("is0 = %lld, is1 = %lld, energy0 = %f, energy1 = %f, %s\n", is0, is1, energy0, energy1, speaker.c_str());
+        }
+
+        // colorful print bug
+        //
+        const char * text = whisper_full_get_segment_text(ctx, i);
+        printf("%s%s", speaker.c_str(), text);
+
+
+        // with timestamps or speakers: each segment on new line
+        if (!params.no_timestamps || params.diarize) {
+            printf("\n");
+        }
+
+        fflush(stdout);
+    }
+}
+
+int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
+    if (params.fname_inp.empty()) {
+        fprintf(stderr, "error: no input files specified\n");
+        return 2;
+    }
+
+    if (params.language != "auto" && whisper_lang_id(params.language.c_str()) == -1) {
+        fprintf(stderr, "error: unknown language '%s'\n", params.language.c_str());
+        exit(0);
+    }
+
+    // whisper init
+
+    struct whisper_context * ctx = whisper_init_from_file(params.model.c_str());
+
+    if (ctx == nullptr) {
+        fprintf(stderr, "error: failed to initialize whisper context\n");
+        return 3;
+    }
+
+    for (int f = 0; f < (int) params.fname_inp.size(); ++f) {
+        const auto fname_inp = params.fname_inp[f];
+        const auto fname_out = f < (int)params.fname_out.size() && !params.fname_out[f].empty() ? params.fname_out[f] : params.fname_inp[f];
+
+        std::vector<float> pcmf32; // mono-channel F32 PCM
+        std::vector<std::vector<float>> pcmf32s; // stereo-channel F32 PCM
+
+        if (!::read_wav(fname_inp, pcmf32, pcmf32s, params.diarize)) {
+            fprintf(stderr, "error: failed to read WAV file '%s'\n", fname_inp.c_str());
+            continue;
+        }
+
+        // print system information
+        {
+            fprintf(stderr, "\n");
+            fprintf(stderr, "system_info: n_threads = %d / %d | %s\n",
+                    params.n_threads*params.n_processors, std::thread::hardware_concurrency(), whisper_print_system_info());
+        }
+
+        // print some info about the processing
+        {
+            fprintf(stderr, "\n");
+            if (!whisper_is_multilingual(ctx)) {
+                if (params.language != "en" || params.translate) {
+                    params.language = "en";
+                    params.translate = false;
+                    fprintf(stderr, "%s: WARNING: model is not multilingual, ignoring language and translation options\n", __func__);
+                }
+            }
+            fprintf(stderr, "%s: processing '%s' (%d samples, %.1f sec), %d threads, %d processors, lang = %s, task = %s, timestamps = %d ...\n",
+                    __func__, fname_inp.c_str(), int(pcmf32.size()), float(pcmf32.size())/WHISPER_SAMPLE_RATE,
+                    params.n_threads, params.n_processors,
+                    params.language.c_str(),
+                    params.translate ? "translate" : "transcribe",
+                    params.no_timestamps ? 0 : 1);
+
+            fprintf(stderr, "\n");
+        }
+
+        // run the inference
+        {
+            whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
+
+            wparams.strategy = params.beam_size > 1 ? WHISPER_SAMPLING_BEAM_SEARCH : WHISPER_SAMPLING_GREEDY;
+
+            wparams.print_realtime   = false;
+            wparams.print_progress   = params.print_progress;
+            wparams.print_timestamps = !params.no_timestamps;
+            wparams.print_special    = params.print_special;
+            wparams.translate        = params.translate;
+            wparams.language         = params.language.c_str();
+            wparams.n_threads        = params.n_threads;
+            wparams.n_max_text_ctx   = params.max_context >= 0 ? params.max_context : wparams.n_max_text_ctx;
+            wparams.offset_ms        = params.offset_t_ms;
+            wparams.duration_ms      = params.duration_ms;
+
+            wparams.token_timestamps = params.output_wts || params.max_len > 0;
+            wparams.thold_pt         = params.word_thold;
+            wparams.entropy_thold    = params.entropy_thold;
+            wparams.logprob_thold    = params.logprob_thold;
+            wparams.max_len          = params.output_wts && params.max_len == 0 ? 60 : params.max_len;
+
+            wparams.speed_up         = params.speed_up;
+
+            wparams.greedy.best_of        = params.best_of;
+            wparams.beam_search.beam_size = params.beam_size;
+
+            wparams.initial_prompt   = params.prompt.c_str();
+
+            whisper_print_user_data user_data = { &params, &pcmf32s };
+
+            // this callback is called on each new segment
+            if (!wparams.print_realtime) {
+                wparams.new_segment_callback           = whisper_print_segment_callback;
+                wparams.new_segment_callback_user_data = &user_data;
+            }
+
+            // example for abort mechanism
+            // in this example, we do not abort the processing, but we could if the flag is set to true
+            // the callback is called before every encoder run - if it returns false, the processing is aborted
+            {
+                static bool is_aborted = false; // NOTE: this should be atomic to avoid data race
+
+                wparams.encoder_begin_callback = [](struct whisper_context * /*ctx*/, struct whisper_state * /*state*/, void * user_data) {
+                    bool is_aborted = *(bool*)user_data;
+                    return !is_aborted;
+                };
+                wparams.encoder_begin_callback_user_data = &is_aborted;
+            }
+
+            if (whisper_full_parallel(ctx, wparams, pcmf32.data(), pcmf32.size(), params.n_processors) != 0) {
+                fprintf(stderr, "failed to process audio\n");
+                return 10;
+            }
+        }
+    }
+
+    const int n_segments = whisper_full_n_segments(ctx);
+    result.resize(n_segments);
+    for (int i = 0; i < n_segments; ++i) {
+        const char * text = whisper_full_get_segment_text(ctx, i);
+        const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
+        const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
+
+        result[i].emplace_back(to_timestamp(t0, true));
+        result[i].emplace_back(to_timestamp(t1, true));
+        result[i].emplace_back(text);
+    }
+
+    whisper_print_timings(ctx);
+    whisper_free(ctx);
+
+    return 0;
+}
+
+class Worker : public Napi::AsyncWorker {
+ public:
+  Worker(Napi::Function& callback, whisper_params params)
+      : Napi::AsyncWorker(callback), params(params) {}
+
+  void Execute() override {
+    run(params, result);
+  }
+
+  void OnOK() override {
+    Napi::HandleScope scope(Env());
+    Napi::Object res = Napi::Array::New(Env(), result.size());
+    for (uint64_t i = 0; i < result.size(); ++i) {
+      Napi::Object tmp = Napi::Array::New(Env(), 3);
+      for (uint64_t j = 0; j < 3; ++j) {
+        tmp[j] = Napi::String::New(Env(), result[i][j]);
+      }
+      res[i] = tmp;
+    }
+    Callback().Call({Env().Null(), res});
+  }
+
+ private:
+  whisper_params params;
+  std::vector<std::vector<std::string>> result;
+};
+
+
+
+Napi::Value whisper(const Napi::CallbackInfo& info) {
+  Napi::Env env = info.Env();
+  if (info.Length() <= 0 || !info[0].IsObject()) {
+    Napi::TypeError::New(env, "object expected").ThrowAsJavaScriptException();
+  }
+  whisper_params params;
+
+  Napi::Object whisper_params = info[0].As<Napi::Object>();
+  std::string language = whisper_params.Get("language").As<Napi::String>();
+  std::string model = whisper_params.Get("model").As<Napi::String>();
+  std::string input = whisper_params.Get("fname_inp").As<Napi::String>();
+
+  params.language = language;
+  params.model = model;
+  params.fname_inp.emplace_back(input);
+
+  Napi::Function callback = info[1].As<Napi::Function>();
+  Worker* worker = new Worker(callback, params);
+  worker->Queue();
+  return env.Undefined();
+}
+
+
+Napi::Object Init(Napi::Env env, Napi::Object exports) {
+  exports.Set(
+      Napi::String::New(env, "whisper"),
+      Napi::Function::New(env, whisper)
+  );
+  return exports;
+}
+
+NODE_API_MODULE(whisper, Init);

examples/addon.node/index.js

@@ -0,0 +1,36 @@
+const path = require("path");
+const { whisper } = require(path.join(
+  __dirname,
+  "../../build/Release/whisper-addon"
+));
+const { promisify } = require("util");
+
+const whisperAsync = promisify(whisper);
+
+const whisperParams = {
+  language: "en",
+  model: path.join(__dirname, "../../models/ggml-base.en.bin"),
+  fname_inp: "../../samples/jfk.wav",
+};
+
+const arguments = process.argv.slice(2);
+const params = Object.fromEntries(
+  arguments.reduce((pre, item) => {
+    if (item.startsWith("--")) {
+      return [...pre, item.slice(2).split("=")];
+    }
+    return pre;
+  }, [])
+);
+
+for (const key in params) {
+  if (whisperParams.hasOwnProperty(key)) {
+    whisperParams[key] = params[key];
+  }
+}
+
+console.log("whisperParams =", whisperParams);
+
+whisperAsync(whisperParams).then((result) => {
+  console.log(`Result from whisper: ${result}`);
+});

examples/addon.node/package.json

@@ -0,0 +1,16 @@
+{
+  "name": "whisper-addon",
+  "version": "0.0.0",
+  "description": "",
+  "main": "index.js",
+  "author": "Qanhe Chen",
+  "license": "MIT",
+  "scripts": {
+    "test": "jest"
+  },
+  "devDependencies": {
+    "cmake-js": "^7.1.1",
+    "jest": "^29.4.0",
+    "node-addon-api": "^5.0.0"
+  }
+}

examples/bench/bench.cpp

@@ -7,6 +7,7 @@
 // command-line parameters
 struct whisper_params {
     int32_t n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency());
+    int32_t what = 0; // what to benchmark: 0 - whisper ecoder, 1 - memcpy, 2 - ggml_mul_mat
 
     std::string model = "models/ggml-base.en.bin";
 };
@@ -23,6 +24,7 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
         }
         else if (arg == "-t" || arg == "--threads") { params.n_threads = std::stoi(argv[++i]); }
         else if (arg == "-m" || arg == "--model")   { params.model     = argv[++i]; }
+        else if (arg == "-w" || arg == "--what")    { params.what     = atoi(argv[++i]); }
         else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             whisper_print_usage(argc, argv, params);
@@ -41,16 +43,14 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -h,       --help        [default] show this help message and exit\n");
     fprintf(stderr, "  -t N,     --threads N   [%-7d] number of threads to use during computation\n", params.n_threads);
     fprintf(stderr, "  -m FNAME, --model FNAME [%-7s] model path\n",                                  params.model.c_str());
+    fprintf(stderr, "  -w N,     --what N      [%-7d] what to benchmark:\n",                          params.what);
+    fprintf(stderr, "                           %-7s  0 - whisper encoder\n",                         "");
+    fprintf(stderr, "                           %-7s  1 - memcpy\n",                                  "");
+    fprintf(stderr, "                           %-7s  2 - ggml_mul_mat\n",                            "");
     fprintf(stderr, "\n");
 }
 
-int main(int argc, char ** argv) {
-    whisper_params params;
-
-    if (whisper_params_parse(argc, argv, params) == false) {
-        return 1;
-    }
-
+int whisper_bench_encoder(const whisper_params & params) {
     // whisper init
 
     struct whisper_context * ctx = whisper_init_from_file(params.model.c_str());
@@ -92,3 +92,22 @@ int main(int argc, char ** argv) {
 
     return 0;
 }
+
+int main(int argc, char ** argv) {
+    whisper_params params;
+
+    if (whisper_params_parse(argc, argv, params) == false) {
+        return 1;
+    }
+
+    int ret = -1;
+
+    switch (params.what) {
+        case 0: ret = whisper_bench_encoder(params);                break;
+        case 1: ret = whisper_bench_memcpy(params.n_threads);       break;
+        case 2: ret = whisper_bench_ggml_mul_mat(params.n_threads); break;
+        default: fprintf(stderr, "error: unknown benchmark: %d\n", params.what); break;
+    }
+
+    return ret;
+}

examples/command.wasm/CMakeLists.txt

@@ -11,6 +11,7 @@ add_executable(${TARGET}
 include(DefaultTargetOptions)
 
 target_link_libraries(${TARGET} PRIVATE
+    common
     whisper
     )
 

examples/command.wasm/emscripten.cpp

@@ -1,4 +1,5 @@
 #include "ggml.h"
+#include "common.h"
 #include "whisper.h"
 
 #include <emscripten.h>
@@ -27,24 +28,6 @@ std::string g_transcribed   = "";
 
 std::vector<float> g_pcmf32;
 
-static std::string trim(const std::string & s) {
-    std::regex e("^\\s+|\\s+$");
-    return std::regex_replace(s, e, "");
-}
-
-static void high_pass_filter(std::vector<float> & data, float cutoff, float sample_rate) {
-    const float rc = 1.0f / (2.0f * M_PI * cutoff);
-    const float dt = 1.0f / sample_rate;
-    const float alpha = dt / (rc + dt);
-
-    float y = data[0];
-
-    for (size_t i = 1; i < data.size(); i++) {
-        y = alpha * (y + data[i] - data[i - 1]);
-        data[i] = y;
-    }
-}
-
 // compute similarity between two strings using Levenshtein distance
 static float similarity(const std::string & s0, const std::string & s1) {
     const size_t len0 = s0.size() + 1;
@@ -75,44 +58,6 @@ void command_set_status(const std::string & status) {
     g_status = status;
 }
 
-bool command_vad_simple(std::vector<float> & pcmf32, int sample_rate, int last_ms, float vad_thold, float freq_thold, bool verbose) {
-    const int n_samples      = pcmf32.size();
-    const int n_samples_last = (sample_rate * last_ms) / 1000;
-
-    if (n_samples_last >= n_samples) {
-        // not enough samples - assume no speech
-        return false;
-    }
-
-    if (freq_thold > 0.0f) {
-        high_pass_filter(pcmf32, freq_thold, sample_rate);
-    }
-
-    float energy_all  = 0.0f;
-    float energy_last = 0.0f;
-
-    for (size_t i = 0; i < n_samples; i++) {
-        energy_all += fabsf(pcmf32[i]);
-
-        if (i >= n_samples - n_samples_last) {
-            energy_last += fabsf(pcmf32[i]);
-        }
-    }
-
-    energy_all  /= n_samples;
-    energy_last /= n_samples_last;
-
-    if (verbose) {
-        fprintf(stderr, "%s: energy_all: %f, energy_last: %f, vad_thold: %f, freq_thold: %f\n", __func__, energy_all, energy_last, vad_thold, freq_thold);
-    }
-
-    if (energy_last > vad_thold*energy_all) {
-        return false;
-    }
-
-    return true;
-}
-
 std::string command_transcribe(whisper_context * ctx, const whisper_full_params & wparams, const std::vector<float> & pcmf32, float & prob, int64_t & t_ms) {
     const auto t_start = std::chrono::high_resolution_clock::now();
 
@@ -155,7 +100,7 @@ void command_get_audio(int ms, int sample_rate, std::vector<float> & audio) {
     const int64_t n_samples = (ms * sample_rate) / 1000;
 
     int64_t n_take = 0;
-    if (g_pcmf32.size() < n_samples) {
+    if (n_samples > (int) g_pcmf32.size()) {
         n_take = g_pcmf32.size();
     } else {
         n_take = n_samples;
@@ -187,7 +132,6 @@ void command_main(size_t index) {
 
     printf("command: using %d threads\n", wparams.n_threads);
 
-    bool is_running   = true;
     bool have_prompt  = false;
     bool ask_prompt   = true;
     bool print_energy = false;
@@ -233,7 +177,7 @@ void command_main(size_t index) {
         {
             command_get_audio(vad_ms, WHISPER_SAMPLE_RATE, pcmf32_cur);
 
-            if (command_vad_simple(pcmf32_cur, WHISPER_SAMPLE_RATE, 1000, vad_thold, freq_thold, print_energy)) {
+            if (::vad_simple(pcmf32_cur, WHISPER_SAMPLE_RATE, 1000, vad_thold, freq_thold, print_energy)) {
                 fprintf(stdout, "%s: Speech detected! Processing ...\n", __func__);
                 command_set_status("Speech detected! Processing ...");
 

examples/command/CMakeLists.txt

@@ -5,6 +5,5 @@ if (WHISPER_SUPPORT_SDL2)
 
     include(DefaultTargetOptions)
 
-    target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS})
-    target_link_libraries(${TARGET} PRIVATE whisper ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
+    target_link_libraries(${TARGET} PRIVATE common common-sdl whisper ${CMAKE_THREAD_LIBS_INIT})
 endif ()

examples/command/command.cpp

@@ -6,11 +6,10 @@
 // ref: https://github.com/ggerganov/whisper.cpp/issues/171
 //
 
+#include "common.h"
+#include "common-sdl.h"
 #include "whisper.h"
 
-#include <SDL.h>
-#include <SDL_audio.h>
-
 #include <sstream>
 #include <cassert>
 #include <cstdio>
@@ -110,309 +109,6 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "\n");
 }
 
-//
-// SDL Audio capture
-//
-
-class audio_async {
-public:
-    audio_async(int len_ms);
-    ~audio_async();
-
-    bool init(int capture_id, int sample_rate);
-
-    // start capturing audio via the provided SDL callback
-    // keep last len_ms seconds of audio in a circular buffer
-    bool resume();
-    bool pause();
-    bool clear();
-
-    // callback to be called by SDL
-    void callback(uint8_t * stream, int len);
-
-    // get audio data from the circular buffer
-    void get(int ms, std::vector<float> & audio);
-
-private:
-    SDL_AudioDeviceID m_dev_id_in = 0;
-
-    int m_len_ms = 0;
-    int m_sample_rate = 0;
-
-    bool       m_running = false;
-    std::mutex m_mutex;
-
-    std::vector<float> m_audio;
-    std::vector<float> m_audio_new;
-    size_t             m_audio_pos = 0;
-    size_t             m_audio_len = 0;
-};
-
-audio_async::audio_async(int len_ms) {
-    m_len_ms = len_ms;
-}
-
-audio_async::~audio_async() {
-    if (m_dev_id_in) {
-        SDL_CloseAudioDevice(m_dev_id_in);
-    }
-}
-
-bool audio_async::init(int capture_id, int sample_rate) {
-    SDL_LogSetPriority(SDL_LOG_CATEGORY_APPLICATION, SDL_LOG_PRIORITY_INFO);
-
-    if (SDL_Init(SDL_INIT_AUDIO) < 0) {
-        SDL_LogError(SDL_LOG_CATEGORY_APPLICATION, "Couldn't initialize SDL: %s\n", SDL_GetError());
-        return false;
-    }
-
-    SDL_SetHintWithPriority(SDL_HINT_AUDIO_RESAMPLING_MODE, "medium", SDL_HINT_OVERRIDE);
-
-    {
-        int nDevices = SDL_GetNumAudioDevices(SDL_TRUE);
-        fprintf(stderr, "%s: found %d capture devices:\n", __func__, nDevices);
-        for (int i = 0; i < nDevices; i++) {
-            fprintf(stderr, "%s:    - Capture device #%d: '%s'\n", __func__, i, SDL_GetAudioDeviceName(i, SDL_TRUE));
-        }
-    }
-
-    SDL_AudioSpec capture_spec_requested;
-    SDL_AudioSpec capture_spec_obtained;
-
-    SDL_zero(capture_spec_requested);
-    SDL_zero(capture_spec_obtained);
-
-    capture_spec_requested.freq     = sample_rate;
-    capture_spec_requested.format   = AUDIO_F32;
-    capture_spec_requested.channels = 1;
-    capture_spec_requested.samples  = 1024;
-    capture_spec_requested.callback = [](void * userdata, uint8_t * stream, int len) {
-        audio_async * audio = (audio_async *) userdata;
-        audio->callback(stream, len);
-    };
-    capture_spec_requested.userdata = this;
-
-    if (capture_id >= 0) {
-        fprintf(stderr, "%s: attempt to open capture device %d : '%s' ...\n", __func__, capture_id, SDL_GetAudioDeviceName(capture_id, SDL_TRUE));
-        m_dev_id_in = SDL_OpenAudioDevice(SDL_GetAudioDeviceName(capture_id, SDL_TRUE), SDL_TRUE, &capture_spec_requested, &capture_spec_obtained, 0);
-    } else {
-        fprintf(stderr, "%s: attempt to open default capture device ...\n", __func__);
-        m_dev_id_in = SDL_OpenAudioDevice(nullptr, SDL_TRUE, &capture_spec_requested, &capture_spec_obtained, 0);
-    }
-
-    if (!m_dev_id_in) {
-        fprintf(stderr, "%s: couldn't open an audio device for capture: %s!\n", __func__, SDL_GetError());
-        m_dev_id_in = 0;
-
-        return false;
-    } else {
-        fprintf(stderr, "%s: obtained spec for input device (SDL Id = %d):\n", __func__, m_dev_id_in);
-        fprintf(stderr, "%s:     - sample rate:       %d\n",                   __func__, capture_spec_obtained.freq);
-        fprintf(stderr, "%s:     - format:            %d (required: %d)\n",    __func__, capture_spec_obtained.format,
-                capture_spec_requested.format);
-        fprintf(stderr, "%s:     - channels:          %d (required: %d)\n",    __func__, capture_spec_obtained.channels,
-                capture_spec_requested.channels);
-        fprintf(stderr, "%s:     - samples per frame: %d\n",                   __func__, capture_spec_obtained.samples);
-    }
-
-    m_sample_rate = capture_spec_obtained.freq;
-
-    m_audio.resize((m_sample_rate*m_len_ms)/1000);
-
-    return true;
-}
-
-bool audio_async::resume() {
-    if (!m_dev_id_in) {
-        fprintf(stderr, "%s: no audio device to resume!\n", __func__);
-        return false;
-    }
-
-    if (m_running) {
-        fprintf(stderr, "%s: already running!\n", __func__);
-        return false;
-    }
-
-    SDL_PauseAudioDevice(m_dev_id_in, 0);
-
-    m_running = true;
-
-    return true;
-}
-
-bool audio_async::pause() {
-    if (!m_dev_id_in) {
-        fprintf(stderr, "%s: no audio device to pause!\n", __func__);
-        return false;
-    }
-
-    if (!m_running) {
-        fprintf(stderr, "%s: already paused!\n", __func__);
-        return false;
-    }
-
-    SDL_PauseAudioDevice(m_dev_id_in, 1);
-
-    m_running = false;
-
-    return true;
-}
-
-bool audio_async::clear() {
-    if (!m_dev_id_in) {
-        fprintf(stderr, "%s: no audio device to clear!\n", __func__);
-        return false;
-    }
-
-    if (!m_running) {
-        fprintf(stderr, "%s: not running!\n", __func__);
-        return false;
-    }
-
-    {
-        std::lock_guard<std::mutex> lock(m_mutex);
-
-        m_audio_pos = 0;
-        m_audio_len = 0;
-    }
-
-    return true;
-}
-
-// callback to be called by SDL
-void audio_async::callback(uint8_t * stream, int len) {
-    if (!m_running) {
-        return;
-    }
-
-    const size_t n_samples = len / sizeof(float);
-
-    m_audio_new.resize(n_samples);
-    memcpy(m_audio_new.data(), stream, n_samples * sizeof(float));
-
-    //fprintf(stderr, "%s: %zu samples, pos %zu, len %zu\n", __func__, n_samples, m_audio_pos, m_audio_len);
-
-    {
-        std::lock_guard<std::mutex> lock(m_mutex);
-
-        if (m_audio_pos + n_samples > m_audio.size()) {
-            const size_t n0 = m_audio.size() - m_audio_pos;
-
-            memcpy(&m_audio[m_audio_pos], stream, n0 * sizeof(float));
-            memcpy(&m_audio[0], &stream[n0], (n_samples - n0) * sizeof(float));
-
-            m_audio_pos = (m_audio_pos + n_samples) % m_audio.size();
-            m_audio_len = m_audio.size();
-        } else {
-            memcpy(&m_audio[m_audio_pos], stream, n_samples * sizeof(float));
-
-            m_audio_pos = (m_audio_pos + n_samples) % m_audio.size();
-            m_audio_len = std::min(m_audio_len + n_samples, m_audio.size());
-        }
-    }
-}
-
-void audio_async::get(int ms, std::vector<float> & result) {
-    if (!m_dev_id_in) {
-        fprintf(stderr, "%s: no audio device to get audio from!\n", __func__);
-        return;
-    }
-
-    if (!m_running) {
-        fprintf(stderr, "%s: not running!\n", __func__);
-        return;
-    }
-
-    result.clear();
-
-    {
-        std::lock_guard<std::mutex> lock(m_mutex);
-
-        if (ms <= 0) {
-            ms = m_len_ms;
-        }
-
-        size_t n_samples = (m_sample_rate * ms) / 1000;
-        if (n_samples > m_audio_len) {
-            n_samples = m_audio_len;
-        }
-
-        result.resize(n_samples);
-
-        int s0 = m_audio_pos - n_samples;
-        if (s0 < 0) {
-            s0 += m_audio.size();
-        }
-
-        if (s0 + n_samples > m_audio.size()) {
-            const size_t n0 = m_audio.size() - s0;
-
-            memcpy(result.data(), &m_audio[s0], n0 * sizeof(float));
-            memcpy(&result[n0], &m_audio[0], (n_samples - n0) * sizeof(float));
-        } else {
-            memcpy(result.data(), &m_audio[s0], n_samples * sizeof(float));
-        }
-    }
-}
-
-///////////////////////////
-
-std::string trim(const std::string & s) {
-    std::regex e("^\\s+|\\s+$");
-    return std::regex_replace(s, e, "");
-}
-
-void high_pass_filter(std::vector<float> & data, float cutoff, float sample_rate) {
-    const float rc = 1.0f / (2.0f * M_PI * cutoff);
-    const float dt = 1.0f / sample_rate;
-    const float alpha = dt / (rc + dt);
-
-    float y = data[0];
-
-    for (size_t i = 1; i < data.size(); i++) {
-        y = alpha * (y + data[i] - data[i - 1]);
-        data[i] = y;
-    }
-}
-
-bool vad_simple(std::vector<float> & pcmf32, int sample_rate, int last_ms, float vad_thold, float freq_thold, bool verbose) {
-    const int n_samples      = pcmf32.size();
-    const int n_samples_last = (sample_rate * last_ms) / 1000;
-
-    if (n_samples_last >= n_samples) {
-        // not enough samples - assume no speech
-        return false;
-    }
-
-    if (freq_thold > 0.0f) {
-        high_pass_filter(pcmf32, freq_thold, sample_rate);
-    }
-
-    float energy_all  = 0.0f;
-    float energy_last = 0.0f;
-
-    for (int i = 0; i < n_samples; i++) {
-        energy_all += fabsf(pcmf32[i]);
-
-        if (i >= n_samples - n_samples_last) {
-            energy_last += fabsf(pcmf32[i]);
-        }
-    }
-
-    energy_all  /= n_samples;
-    energy_last /= n_samples_last;
-
-    if (verbose) {
-        fprintf(stderr, "%s: energy_all: %f, energy_last: %f, vad_thold: %f, freq_thold: %f\n", __func__, energy_all, energy_last, vad_thold, freq_thold);
-    }
-
-    if (energy_last > vad_thold*energy_all) {
-        return false;
-    }
-
-    return true;
-}
-
 std::string transcribe(whisper_context * ctx, const whisper_params & params, const std::vector<float> & pcmf32, float & prob, int64_t & t_ms) {
     const auto t_start = std::chrono::high_resolution_clock::now();
 
@@ -502,7 +198,7 @@ std::vector<std::string> read_allowed_commands(const std::string & fname) {
 
     std::string line;
     while (std::getline(ifs, line)) {
-        line = trim(line);
+        line = ::trim(line);
         if (line.empty()) {
             continue;
         }
@@ -526,23 +222,6 @@ std::vector<std::string> get_words(const std::string &txt) {
     return words;
 }
 
-// returns true if no exit event was received
-bool process_sdl_events() {
-    SDL_Event event;
-    while (SDL_PollEvent(&event)) {
-        switch (event.type) {
-            case SDL_QUIT:
-                {
-                    return false;
-                } break;
-            default:
-                break;
-        }
-    }
-
-    return true;
-}
-
 // command-list mode
 // guide the transcription to match the most likely command from a provided list
 int process_command_list(struct whisper_context * ctx, audio_async &audio, const whisper_params &params) {
@@ -634,14 +313,14 @@ int process_command_list(struct whisper_context * ctx, audio_async &audio, const
     // main loop
     while (is_running) {
         // handle Ctrl + C
-        is_running = process_sdl_events();
+        is_running = sdl_poll_events();
 
         // delay
         std::this_thread::sleep_for(std::chrono::milliseconds(100));
 
         audio.get(2000, pcmf32_cur);
 
-        if (vad_simple(pcmf32_cur, WHISPER_SAMPLE_RATE, 1000, params.vad_thold, params.freq_thold, params.print_energy)) {
+        if (::vad_simple(pcmf32_cur, WHISPER_SAMPLE_RATE, 1000, params.vad_thold, params.freq_thold, params.print_energy)) {
             fprintf(stdout, "%s: Speech detected! Processing ...\n", __func__);
 
             const auto t_start = std::chrono::high_resolution_clock::now();
@@ -775,7 +454,7 @@ int always_prompt_transcription(struct whisper_context * ctx, audio_async & audi
     // main loop
     while (is_running) {
         // handle Ctrl + C
-        is_running = process_sdl_events();
+        is_running = sdl_poll_events();
 
         // delay
         std::this_thread::sleep_for(std::chrono::milliseconds(100));
@@ -791,7 +470,7 @@ int always_prompt_transcription(struct whisper_context * ctx, audio_async & audi
         {
             audio.get(2000, pcmf32_cur);
 
-            if (vad_simple(pcmf32_cur, WHISPER_SAMPLE_RATE, 1000, params.vad_thold, params.freq_thold, params.print_energy)) {
+            if (::vad_simple(pcmf32_cur, WHISPER_SAMPLE_RATE, 1000, params.vad_thold, params.freq_thold, params.print_energy)) {
                 fprintf(stdout, "%s: Speech detected! Processing ...\n", __func__);
 
                 int64_t t_ms = 0;
@@ -854,7 +533,7 @@ int process_general_transcription(struct whisper_context * ctx, audio_async &aud
     // main loop
     while (is_running) {
         // handle Ctrl + C
-        is_running = process_sdl_events();
+        is_running = sdl_poll_events();
 
         // delay
         std::this_thread::sleep_for(std::chrono::milliseconds(100));
@@ -870,7 +549,7 @@ int process_general_transcription(struct whisper_context * ctx, audio_async &aud
         {
             audio.get(2000, pcmf32_cur);
 
-            if (vad_simple(pcmf32_cur, WHISPER_SAMPLE_RATE, 1000, params.vad_thold, params.freq_thold, params.print_energy)) {
+            if (::vad_simple(pcmf32_cur, WHISPER_SAMPLE_RATE, 1000, params.vad_thold, params.freq_thold, params.print_energy)) {
                 fprintf(stdout, "%s: Speech detected! Processing ...\n", __func__);
 
                 int64_t t_ms = 0;

examples/common-sdl.cpp

@@ -0,0 +1,226 @@
+#include "common-sdl.h"
+
+audio_async::audio_async(int len_ms) {
+    m_len_ms = len_ms;
+
+    m_running = false;
+}
+
+audio_async::~audio_async() {
+    if (m_dev_id_in) {
+        SDL_CloseAudioDevice(m_dev_id_in);
+    }
+}
+
+bool audio_async::init(int capture_id, int sample_rate) {
+    SDL_LogSetPriority(SDL_LOG_CATEGORY_APPLICATION, SDL_LOG_PRIORITY_INFO);
+
+    if (SDL_Init(SDL_INIT_AUDIO) < 0) {
+        SDL_LogError(SDL_LOG_CATEGORY_APPLICATION, "Couldn't initialize SDL: %s\n", SDL_GetError());
+        return false;
+    }
+
+    SDL_SetHintWithPriority(SDL_HINT_AUDIO_RESAMPLING_MODE, "medium", SDL_HINT_OVERRIDE);
+
+    {
+        int nDevices = SDL_GetNumAudioDevices(SDL_TRUE);
+        fprintf(stderr, "%s: found %d capture devices:\n", __func__, nDevices);
+        for (int i = 0; i < nDevices; i++) {
+            fprintf(stderr, "%s:    - Capture device #%d: '%s'\n", __func__, i, SDL_GetAudioDeviceName(i, SDL_TRUE));
+        }
+    }
+
+    SDL_AudioSpec capture_spec_requested;
+    SDL_AudioSpec capture_spec_obtained;
+
+    SDL_zero(capture_spec_requested);
+    SDL_zero(capture_spec_obtained);
+
+    capture_spec_requested.freq     = sample_rate;
+    capture_spec_requested.format   = AUDIO_F32;
+    capture_spec_requested.channels = 1;
+    capture_spec_requested.samples  = 1024;
+    capture_spec_requested.callback = [](void * userdata, uint8_t * stream, int len) {
+        audio_async * audio = (audio_async *) userdata;
+        audio->callback(stream, len);
+    };
+    capture_spec_requested.userdata = this;
+
+    if (capture_id >= 0) {
+        fprintf(stderr, "%s: attempt to open capture device %d : '%s' ...\n", __func__, capture_id, SDL_GetAudioDeviceName(capture_id, SDL_TRUE));
+        m_dev_id_in = SDL_OpenAudioDevice(SDL_GetAudioDeviceName(capture_id, SDL_TRUE), SDL_TRUE, &capture_spec_requested, &capture_spec_obtained, 0);
+    } else {
+        fprintf(stderr, "%s: attempt to open default capture device ...\n", __func__);
+        m_dev_id_in = SDL_OpenAudioDevice(nullptr, SDL_TRUE, &capture_spec_requested, &capture_spec_obtained, 0);
+    }
+
+    if (!m_dev_id_in) {
+        fprintf(stderr, "%s: couldn't open an audio device for capture: %s!\n", __func__, SDL_GetError());
+        m_dev_id_in = 0;
+
+        return false;
+    } else {
+        fprintf(stderr, "%s: obtained spec for input device (SDL Id = %d):\n", __func__, m_dev_id_in);
+        fprintf(stderr, "%s:     - sample rate:       %d\n",                   __func__, capture_spec_obtained.freq);
+        fprintf(stderr, "%s:     - format:            %d (required: %d)\n",    __func__, capture_spec_obtained.format,
+                capture_spec_requested.format);
+        fprintf(stderr, "%s:     - channels:          %d (required: %d)\n",    __func__, capture_spec_obtained.channels,
+                capture_spec_requested.channels);
+        fprintf(stderr, "%s:     - samples per frame: %d\n",                   __func__, capture_spec_obtained.samples);
+    }
+
+    m_sample_rate = capture_spec_obtained.freq;
+
+    m_audio.resize((m_sample_rate*m_len_ms)/1000);
+
+    return true;
+}
+
+bool audio_async::resume() {
+    if (!m_dev_id_in) {
+        fprintf(stderr, "%s: no audio device to resume!\n", __func__);
+        return false;
+    }
+
+    if (m_running) {
+        fprintf(stderr, "%s: already running!\n", __func__);
+        return false;
+    }
+
+    SDL_PauseAudioDevice(m_dev_id_in, 0);
+
+    m_running = true;
+
+    return true;
+}
+
+bool audio_async::pause() {
+    if (!m_dev_id_in) {
+        fprintf(stderr, "%s: no audio device to pause!\n", __func__);
+        return false;
+    }
+
+    if (!m_running) {
+        fprintf(stderr, "%s: already paused!\n", __func__);
+        return false;
+    }
+
+    SDL_PauseAudioDevice(m_dev_id_in, 1);
+
+    m_running = false;
+
+    return true;
+}
+
+bool audio_async::clear() {
+    if (!m_dev_id_in) {
+        fprintf(stderr, "%s: no audio device to clear!\n", __func__);
+        return false;
+    }
+
+    if (!m_running) {
+        fprintf(stderr, "%s: not running!\n", __func__);
+        return false;
+    }
+
+    {
+        std::lock_guard<std::mutex> lock(m_mutex);
+
+        m_audio_pos = 0;
+        m_audio_len = 0;
+    }
+
+    return true;
+}
+
+// callback to be called by SDL
+void audio_async::callback(uint8_t * stream, int len) {
+    if (!m_running) {
+        return;
+    }
+
+    const size_t n_samples = len / sizeof(float);
+
+    m_audio_new.resize(n_samples);
+    memcpy(m_audio_new.data(), stream, n_samples * sizeof(float));
+
+    //fprintf(stderr, "%s: %zu samples, pos %zu, len %zu\n", __func__, n_samples, m_audio_pos, m_audio_len);
+
+    {
+        std::lock_guard<std::mutex> lock(m_mutex);
+
+        if (m_audio_pos + n_samples > m_audio.size()) {
+            const size_t n0 = m_audio.size() - m_audio_pos;
+
+            memcpy(&m_audio[m_audio_pos], stream, n0 * sizeof(float));
+            memcpy(&m_audio[0], &stream[n0], (n_samples - n0) * sizeof(float));
+
+            m_audio_pos = (m_audio_pos + n_samples) % m_audio.size();
+            m_audio_len = m_audio.size();
+        } else {
+            memcpy(&m_audio[m_audio_pos], stream, n_samples * sizeof(float));
+
+            m_audio_pos = (m_audio_pos + n_samples) % m_audio.size();
+            m_audio_len = std::min(m_audio_len + n_samples, m_audio.size());
+        }
+    }
+}
+
+void audio_async::get(int ms, std::vector<float> & result) {
+    if (!m_dev_id_in) {
+        fprintf(stderr, "%s: no audio device to get audio from!\n", __func__);
+        return;
+    }
+
+    if (!m_running) {
+        fprintf(stderr, "%s: not running!\n", __func__);
+        return;
+    }
+
+    result.clear();
+
+    {
+        std::lock_guard<std::mutex> lock(m_mutex);
+
+        if (ms <= 0) {
+            ms = m_len_ms;
+        }
+
+        size_t n_samples = (m_sample_rate * ms) / 1000;
+        if (n_samples > m_audio_len) {
+            n_samples = m_audio_len;
+        }
+
+        result.resize(n_samples);
+
+        int s0 = m_audio_pos - n_samples;
+        if (s0 < 0) {
+            s0 += m_audio.size();
+        }
+
+        if (s0 + n_samples > m_audio.size()) {
+            const size_t n0 = m_audio.size() - s0;
+
+            memcpy(result.data(), &m_audio[s0], n0 * sizeof(float));
+            memcpy(&result[n0], &m_audio[0], (n_samples - n0) * sizeof(float));
+        } else {
+            memcpy(result.data(), &m_audio[s0], n_samples * sizeof(float));
+        }
+    }
+}
+
+bool sdl_poll_events() {
+    SDL_Event event;
+    while (SDL_PollEvent(&event)) {
+        switch (event.type) {
+            case SDL_QUIT:
+                {
+                    return false;
+                } break;
+            default:
+                break;
+        }
+    }
+
+    return true;
+}

examples/common-sdl.h

@@ -0,0 +1,50 @@
+#pragma once
+
+#include <SDL.h>
+#include <SDL_audio.h>
+
+#include <atomic>
+#include <cstdint>
+#include <vector>
+#include <mutex>
+
+//
+// SDL Audio capture
+//
+
+class audio_async {
+public:
+    audio_async(int len_ms);
+    ~audio_async();
+
+    bool init(int capture_id, int sample_rate);
+
+    // start capturing audio via the provided SDL callback
+    // keep last len_ms seconds of audio in a circular buffer
+    bool resume();
+    bool pause();
+    bool clear();
+
+    // callback to be called by SDL
+    void callback(uint8_t * stream, int len);
+
+    // get audio data from the circular buffer
+    void get(int ms, std::vector<float> & audio);
+
+private:
+    SDL_AudioDeviceID m_dev_id_in = 0;
+
+    int m_len_ms = 0;
+    int m_sample_rate = 0;
+
+    std::atomic_bool m_running;
+    std::mutex       m_mutex;
+
+    std::vector<float> m_audio;
+    std::vector<float> m_audio_new;
+    size_t             m_audio_pos = 0;
+    size_t             m_audio_len = 0;
+};
+
+// Return false if need to quit
+bool sdl_poll_events();

examples/common.cpp

@@ -0,0 +1,162 @@
+#include "common.h"
+
+// third-party utilities
+// use your favorite implementations
+#define DR_WAV_IMPLEMENTATION
+#include "dr_wav.h"
+
+#include <cmath>
+#include <regex>
+
+#ifndef M_PI
+#define M_PI 3.14159265358979323846
+#endif
+
+std::string trim(const std::string & s) {
+    std::regex e("^\\s+|\\s+$");
+    return std::regex_replace(s, e, "");
+}
+
+std::string replace(const std::string & s, const std::string & from, const std::string & to) {
+    std::string result = s;
+    size_t pos = 0;
+    while ((pos = result.find(from, pos)) != std::string::npos) {
+        result.replace(pos, from.length(), to);
+        pos += to.length();
+    }
+    return result;
+}
+
+bool read_wav(const std::string & fname, std::vector<float>& pcmf32, std::vector<std::vector<float>>& pcmf32s, bool stereo) {
+    drwav wav;
+    std::vector<uint8_t> wav_data; // used for pipe input from stdin
+
+    if (fname == "-") {
+        {
+            uint8_t buf[1024];
+            while (true)
+            {
+                const size_t n = fread(buf, 1, sizeof(buf), stdin);
+                if (n == 0) {
+                    break;
+                }
+                wav_data.insert(wav_data.end(), buf, buf + n);
+            }
+        }
+
+        if (drwav_init_memory(&wav, wav_data.data(), wav_data.size(), nullptr) == false) {
+            fprintf(stderr, "error: failed to open WAV file from stdin\n");
+            return false;
+        }
+
+        fprintf(stderr, "%s: read %zu bytes from stdin\n", __func__, wav_data.size());
+    }
+    else if (drwav_init_file(&wav, fname.c_str(), nullptr) == false) {
+        fprintf(stderr, "error: failed to open '%s' as WAV file\n", fname.c_str());
+        return false;
+    }
+
+    if (wav.channels != 1 && wav.channels != 2) {
+        fprintf(stderr, "%s: WAV file '%s' must be mono or stereo\n", __func__, fname.c_str());
+        return false;
+    }
+
+    if (stereo && wav.channels != 2) {
+        fprintf(stderr, "%s: WAV file '%s' must be stereo for diarization\n", __func__, fname.c_str());
+        return false;
+    }
+
+    if (wav.sampleRate != COMMON_SAMPLE_RATE) {
+        fprintf(stderr, "%s: WAV file '%s' must be %i kHz\n", __func__, fname.c_str(), COMMON_SAMPLE_RATE/1000);
+        return false;
+    }
+
+    if (wav.bitsPerSample != 16) {
+        fprintf(stderr, "%s: WAV file '%s' must be 16-bit\n", __func__, fname.c_str());
+        return false;
+    }
+
+    const uint64_t n = wav_data.empty() ? wav.totalPCMFrameCount : wav_data.size()/(wav.channels*wav.bitsPerSample/8);
+
+    std::vector<int16_t> pcm16;
+    pcm16.resize(n*wav.channels);
+    drwav_read_pcm_frames_s16(&wav, n, pcm16.data());
+    drwav_uninit(&wav);
+
+    // convert to mono, float
+    pcmf32.resize(n);
+    if (wav.channels == 1) {
+        for (uint64_t i = 0; i < n; i++) {
+            pcmf32[i] = float(pcm16[i])/32768.0f;
+        }
+    } else {
+        for (uint64_t i = 0; i < n; i++) {
+            pcmf32[i] = float(pcm16[2*i] + pcm16[2*i + 1])/65536.0f;
+        }
+    }
+
+    if (stereo) {
+        // convert to stereo, float
+        pcmf32s.resize(2);
+
+        pcmf32s[0].resize(n);
+        pcmf32s[1].resize(n);
+        for (uint64_t i = 0; i < n; i++) {
+            pcmf32s[0][i] = float(pcm16[2*i])/32768.0f;
+            pcmf32s[1][i] = float(pcm16[2*i + 1])/32768.0f;
+        }
+    }
+
+    return true;
+}
+
+void high_pass_filter(std::vector<float> & data, float cutoff, float sample_rate) {
+    const float rc = 1.0f / (2.0f * M_PI * cutoff);
+    const float dt = 1.0f / sample_rate;
+    const float alpha = dt / (rc + dt);
+
+    float y = data[0];
+
+    for (size_t i = 1; i < data.size(); i++) {
+        y = alpha * (y + data[i] - data[i - 1]);
+        data[i] = y;
+    }
+}
+
+bool vad_simple(std::vector<float> & pcmf32, int sample_rate, int last_ms, float vad_thold, float freq_thold, bool verbose) {
+    const int n_samples      = pcmf32.size();
+    const int n_samples_last = (sample_rate * last_ms) / 1000;
+
+    if (n_samples_last >= n_samples) {
+        // not enough samples - assume no speech
+        return false;
+    }
+
+    if (freq_thold > 0.0f) {
+        high_pass_filter(pcmf32, freq_thold, sample_rate);
+    }
+
+    float energy_all  = 0.0f;
+    float energy_last = 0.0f;
+
+    for (int i = 0; i < n_samples; i++) {
+        energy_all += fabsf(pcmf32[i]);
+
+        if (i >= n_samples - n_samples_last) {
+            energy_last += fabsf(pcmf32[i]);
+        }
+    }
+
+    energy_all  /= n_samples;
+    energy_last /= n_samples_last;
+
+    if (verbose) {
+        fprintf(stderr, "%s: energy_all: %f, energy_last: %f, vad_thold: %f, freq_thold: %f\n", __func__, energy_all, energy_last, vad_thold, freq_thold);
+    }
+
+    if (energy_last > vad_thold*energy_all) {
+        return false;
+    }
+
+    return true;
+}

examples/common.h

@@ -0,0 +1,40 @@
+#pragma once
+
+// needs to match WHISPER_SAMPLE_RATE
+#define COMMON_SAMPLE_RATE 16000
+
+#include <vector>
+#include <string>
+
+std::string trim(const std::string & s);
+
+std::string replace(
+        const std::string & s,
+        const std::string & from,
+        const std::string & to);
+
+// Read WAV audio file and store the PCM data into pcmf32
+// The sample rate of the audio must be equal to COMMON_SAMPLE_RATE
+// If stereo flag is set and the audio has 2 channels, the pcmf32s will contain 2 channel PCM
+bool read_wav(
+        const std::string & fname,
+        std::vector<float> & pcmf32,
+        std::vector<std::vector<float>> & pcmf32s,
+        bool stereo);
+
+// Apply a high-pass frequency filter to PCM audio
+// Suppresses frequencies below cutoff Hz
+void high_pass_filter(
+        std::vector<float> & data,
+        float cutoff,
+        float sample_rate);
+
+// Basic voice activity detection (VAD) using audio energy adaptive threshold
+bool vad_simple(
+        std::vector<float> & pcmf32,
+        int   sample_rate,
+        int   last_ms,
+        float vad_thold,
+        float freq_thold,
+        bool  verbose);
+

examples/helpers.js

@@ -8,7 +8,7 @@ function convertTypedArray(src, type) {
 
 var printTextarea = (function() {
     var element = document.getElementById('output');
-    if (element) element.alue = ''; // clear browser cache
+    if (element) element.value = ''; // clear browser cache
     return function(text) {
         if (arguments.length > 1) text = Array.prototype.slice.call(arguments).join(' ');
         console.log(text);
@@ -88,11 +88,15 @@ async function fetchRemote(url, cbProgress, cbPrint) {
 // - check if the data is already in the IndexedDB
 // - if not, fetch it from the remote URL and store it in the IndexedDB
 function loadRemote(url, dst, size_mb, cbProgress, cbReady, cbCancel, cbPrint) {
-    // query the storage quota and print it
-    navigator.storage.estimate().then(function (estimate) {
-        cbPrint('loadRemote: storage quota: ' + estimate.quota + ' bytes');
-        cbPrint('loadRemote: storage usage: ' + estimate.usage + ' bytes');
-    });
+    if (!navigator.storage || !navigator.storage.estimate) {
+        cbPrint('loadRemote: navigator.storage.estimate() is not supported');
+    } else {
+        // query the storage quota and print it
+        navigator.storage.estimate().then(function (estimate) {
+            cbPrint('loadRemote: storage quota: ' + estimate.quota + ' bytes');
+            cbPrint('loadRemote: storage usage: ' + estimate.usage + ' bytes');
+        });
+    }
 
     // check if the data is already in the IndexedDB
     var rq = indexedDB.open(dbName, dbVersion);

examples/livestream.sh

@@ -100,7 +100,7 @@ while [ $running -eq 1 ]; do
         err=$(cat /tmp/whisper-live.err | wc -l)
     done
 
-    ./main -t 8 -m ./models/ggml-base.en.bin -f /tmp/whisper-live.wav --no-timestamps -otxt 2> /tmp/whispererr | tail -n 1
+    ./main -t 8 -m ./models/ggml-${model}.bin -f /tmp/whisper-live.wav --no-timestamps -otxt 2> /tmp/whispererr | tail -n 1
 
     while [ $SECONDS -lt $((($i+1)*$step_s)) ]; do
         sleep 1

examples/main/CMakeLists.txt

@@ -3,4 +3,4 @@ add_executable(${TARGET} main.cpp)
 
 include(DefaultTargetOptions)
 
-target_link_libraries(${TARGET} PRIVATE whisper ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE common whisper ${CMAKE_THREAD_LIBS_INIT})

examples/main/README.md

@@ -9,25 +9,36 @@ It can be used as a reference for using the `whisper.cpp` library in other proje
 usage: ./main [options] file0.wav file1.wav ...
 
 options:
-  -h,       --help          [default] show this help message and exit
-  -t N,     --threads N     [4      ] number of threads to use during computation
-  -p N,     --processors N  [1      ] number of processors to use during computation
-  -ot N,    --offset-t N    [0      ] time offset in milliseconds
-  -on N,    --offset-n N    [0      ] segment index offset
-  -d  N,    --duration N    [0      ] duration of audio to process in milliseconds
-  -mc N,    --max-context N [-1     ] maximum number of text context tokens to store
-  -ml N,    --max-len N     [0      ] maximum segment length in characters
-  -wt N,    --word-thold N  [0.01   ] word timestamp probability threshold
-  -su,      --speed-up      [false  ] speed up audio by x2 (reduced accuracy)
-  -tr,      --translate     [false  ] translate from source language to english
-  -otxt,    --output-txt    [false  ] output result in a text file
-  -ovtt,    --output-vtt    [false  ] output result in a vtt file
-  -osrt,    --output-srt    [false  ] output result in a srt file
-  -owts,    --output-words  [false  ] output script for generating karaoke video
-  -ps,      --print-special [false  ] print special tokens
-  -pc,      --print-colors  [false  ] print colors
-  -nt,      --no-timestamps [true   ] do not print timestamps
-  -l LANG,  --language LANG [en     ] spoken language
-  -m FNAME, --model FNAME   [models/ggml-base.en.bin] model path
-  -f FNAME, --file FNAME    [       ] input WAV file path
+  -h,        --help              [default] show this help message and exit
+  -t N,      --threads N         [4      ] number of threads to use during computation
+  -p N,      --processors N      [1      ] number of processors to use during computation
+  -ot N,     --offset-t N        [0      ] time offset in milliseconds
+  -on N,     --offset-n N        [0      ] segment index offset
+  -d  N,     --duration N        [0      ] duration of audio to process in milliseconds
+  -mc N,     --max-context N     [-1     ] maximum number of text context tokens to store
+  -ml N,     --max-len N         [0      ] maximum segment length in characters
+  -bo N,     --best-of N         [5      ] number of best candidates to keep
+  -bs N,     --beam-size N       [-1     ] beam size for beam search
+  -wt N,     --word-thold N      [0.01   ] word timestamp probability threshold
+  -et N,     --entropy-thold N   [2.40   ] entropy threshold for decoder fail
+  -lpt N,    --logprob-thold N   [-1.00  ] log probability threshold for decoder fail
+  -su,       --speed-up          [false  ] speed up audio by x2 (reduced accuracy)
+  -tr,       --translate         [false  ] translate from source language to english
+  -di,       --diarize           [false  ] stereo audio diarization
+  -nf,       --no-fallback       [false  ] do not use temperature fallback while decoding
+  -otxt,     --output-txt        [false  ] output result in a text file
+  -ovtt,     --output-vtt        [false  ] output result in a vtt file
+  -osrt,     --output-srt        [false  ] output result in a srt file
+  -owts,     --output-words      [false  ] output script for generating karaoke video
+  -ocsv,     --output-csv        [false  ] output result in a CSV file
+  -oj,       --output-json       [false  ] output result in a JSON file
+  -of FNAME, --output-file FNAME [       ] output file path (without file extension)
+  -ps,       --print-special     [false  ] print special tokens
+  -pc,       --print-colors      [false  ] print colors
+  -pp,       --print-progress    [false  ] print progress
+  -nt,       --no-timestamps     [true   ] do not print timestamps
+  -l LANG,   --language LANG     [en     ] spoken language ('auto' for auto-detect)
+             --prompt PROMPT     [       ] initial prompt
+  -m FNAME,  --model FNAME       [models/ggml-base.en.bin] model path
+  -f FNAME,  --file FNAME        [       ] input WAV file path
 ```

examples/main/main.cpp

@@ -1,9 +1,6 @@
-#include "whisper.h"
+#include "common.h"
 
-// third-party utilities
-// use your favorite implementations
-#define DR_WAV_IMPLEMENTATION
-#include "dr_wav.h"
+#include "whisper.h"
 
 #include <cmath>
 #include <fstream>
@@ -11,6 +8,7 @@
 #include <string>
 #include <thread>
 #include <vector>
+#include <cstring>
 
 // Terminal color map. 10 colors grouped in ranges [0.0, 0.1, ..., 0.9]
 // Lowest is red, middle is yellow, highest is green.
@@ -53,27 +51,31 @@ void replace_all(std::string & s, const std::string & search, const std::string
 // command-line parameters
 struct whisper_params {
     int32_t n_threads    = std::min(4, (int32_t) std::thread::hardware_concurrency());
-    int32_t n_processors = 1;
-    int32_t offset_t_ms  = 0;
-    int32_t offset_n     = 0;
-    int32_t duration_ms  = 0;
+    int32_t n_processors =  1;
+    int32_t offset_t_ms  =  0;
+    int32_t offset_n     =  0;
+    int32_t duration_ms  =  0;
     int32_t max_context  = -1;
-    int32_t max_len      = 0;
-    int32_t best_of      = 5;
+    int32_t max_len      =  0;
+    int32_t best_of      =  2;
     int32_t beam_size    = -1;
 
-    float word_thold    = 0.01f;
-    float entropy_thold = 2.4f;
-    float logprob_thold = -1.0f;
+    float word_thold    =  0.01f;
+    float entropy_thold =  2.40f;
+    float logprob_thold = -1.00f;
 
     bool speed_up       = false;
     bool translate      = false;
     bool diarize        = false;
+    bool split_on_word  = false;
+    bool no_fallback    = false;
     bool output_txt     = false;
     bool output_vtt     = false;
     bool output_srt     = false;
     bool output_wts     = false;
     bool output_csv     = false;
+    bool output_jsn     = false;
+    bool output_lrc     = false;
     bool print_special  = false;
     bool print_colors   = false;
     bool print_progress = false;
@@ -81,9 +83,11 @@ struct whisper_params {
 
     std::string language = "en";
     std::string prompt;
+    std::string font_path = "/System/Library/Fonts/Supplemental/Courier New Bold.ttf";
     std::string model    = "models/ggml-base.en.bin";
 
     std::vector<std::string> fname_inp = {};
+    std::vector<std::string> fname_out = {};
 };
 
 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
@@ -92,6 +96,11 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
     for (int i = 1; i < argc; i++) {
         std::string arg = argv[i];
 
+        if (arg == "-"){
+            params.fname_inp.push_back(arg);
+            continue;
+        }
+
         if (arg[0] != '-') {
             params.fname_inp.push_back(arg);
             continue;
@@ -116,11 +125,17 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
         else if (arg == "-su"   || arg == "--speed-up")       { params.speed_up       = true; }
         else if (arg == "-tr"   || arg == "--translate")      { params.translate      = true; }
         else if (arg == "-di"   || arg == "--diarize")        { params.diarize        = true; }
+        else if (arg == "-sow"  || arg == "--split-on-word")  { params.split_on_word  = true; }
+        else if (arg == "-nf"   || arg == "--no-fallback")    { params.no_fallback    = true; }
         else if (arg == "-otxt" || arg == "--output-txt")     { params.output_txt     = true; }
         else if (arg == "-ovtt" || arg == "--output-vtt")     { params.output_vtt     = true; }
         else if (arg == "-osrt" || arg == "--output-srt")     { params.output_srt     = true; }
         else if (arg == "-owts" || arg == "--output-words")   { params.output_wts     = true; }
+        else if (arg == "-olrc" || arg == "--output-lrc")     { params.output_lrc     = true; }
+        else if (arg == "-fp"   || arg == "--font-path")      { params.font_path      = argv[++i]; }
         else if (arg == "-ocsv" || arg == "--output-csv")     { params.output_csv     = true; }
+        else if (arg == "-oj"   || arg == "--output-json")    { params.output_jsn     = true; }
+        else if (arg == "-of"   || arg == "--output-file")    { params.fname_out.emplace_back(argv[++i]); }
         else if (arg == "-ps"   || arg == "--print-special")  { params.print_special  = true; }
         else if (arg == "-pc"   || arg == "--print-colors")   { params.print_colors   = true; }
         else if (arg == "-pp"   || arg == "--print-progress") { params.print_progress = true; }
@@ -144,35 +159,41 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "usage: %s [options] file0.wav file1.wav ...\n", argv[0]);
     fprintf(stderr, "\n");
     fprintf(stderr, "options:\n");
-    fprintf(stderr, "  -h,       --help            [default] show this help message and exit\n");
-    fprintf(stderr, "  -t N,     --threads N       [%-7d] number of threads to use during computation\n",    params.n_threads);
-    fprintf(stderr, "  -p N,     --processors N    [%-7d] number of processors to use during computation\n", params.n_processors);
-    fprintf(stderr, "  -ot N,    --offset-t N      [%-7d] time offset in milliseconds\n",                    params.offset_t_ms);
-    fprintf(stderr, "  -on N,    --offset-n N      [%-7d] segment index offset\n",                           params.offset_n);
-    fprintf(stderr, "  -d  N,    --duration N      [%-7d] duration of audio to process in milliseconds\n",   params.duration_ms);
-    fprintf(stderr, "  -mc N,    --max-context N   [%-7d] maximum number of text context tokens to store\n", params.max_context);
-    fprintf(stderr, "  -ml N,    --max-len N       [%-7d] maximum segment length in characters\n",           params.max_len);
-    fprintf(stderr, "  -bo N,    --best-of N       [%-7d] number of best candidates to keep\n",              params.best_of);
-    fprintf(stderr, "  -bs N,    --beam-size N     [%-7d] beam size for beam search\n",                      params.beam_size);
-    fprintf(stderr, "  -wt N,    --word-thold N    [%-7.2f] word timestamp probability threshold\n",         params.word_thold);
-    fprintf(stderr, "  -et N,    --entropy-thold N [%-7.2f] entropy threshold for decoder fail\n",           params.entropy_thold);
-    fprintf(stderr, "  -lpt N,   --logprob-thold N [%-7.2f] log probability threshold for decoder fail\n",   params.logprob_thold);
-    fprintf(stderr, "  -su,      --speed-up        [%-7s] speed up audio by x2 (reduced accuracy)\n",        params.speed_up ? "true" : "false");
-    fprintf(stderr, "  -tr,      --translate       [%-7s] translate from source language to english\n",      params.translate ? "true" : "false");
-    fprintf(stderr, "  -di,      --diarize         [%-7s] stereo audio diarization\n",                       params.diarize ? "true" : "false");
-    fprintf(stderr, "  -otxt,    --output-txt      [%-7s] output result in a text file\n",                   params.output_txt ? "true" : "false");
-    fprintf(stderr, "  -ovtt,    --output-vtt      [%-7s] output result in a vtt file\n",                    params.output_vtt ? "true" : "false");
-    fprintf(stderr, "  -osrt,    --output-srt      [%-7s] output result in a srt file\n",                    params.output_srt ? "true" : "false");
-    fprintf(stderr, "  -owts,    --output-words    [%-7s] output script for generating karaoke video\n",     params.output_wts ? "true" : "false");
-    fprintf(stderr, "  -ocsv,    --output-csv      [%-7s] output result in a CSV file\n",                    params.output_csv ? "true" : "false");
-    fprintf(stderr, "  -ps,      --print-special   [%-7s] print special tokens\n",                           params.print_special ? "true" : "false");
-    fprintf(stderr, "  -pc,      --print-colors    [%-7s] print colors\n",                                   params.print_colors ? "true" : "false");
-    fprintf(stderr, "  -pp,      --print-progress  [%-7s] print progress\n",                                 params.print_progress ? "true" : "false");
-    fprintf(stderr, "  -nt,      --no-timestamps   [%-7s] do not print timestamps\n",                        params.no_timestamps ? "false" : "true");
-    fprintf(stderr, "  -l LANG,  --language LANG   [%-7s] spoken language ('auto' for auto-detect)\n",       params.language.c_str());
-    fprintf(stderr, "            --prompt PROMPT   [%-7s] initial prompt\n",                                 params.prompt.c_str());
-    fprintf(stderr, "  -m FNAME, --model FNAME     [%-7s] model path\n",                                     params.model.c_str());
-    fprintf(stderr, "  -f FNAME, --file FNAME      [%-7s] input WAV file path\n",                            "");
+    fprintf(stderr, "  -h,        --help              [default] show this help message and exit\n");
+    fprintf(stderr, "  -t N,      --threads N         [%-7d] number of threads to use during computation\n",    params.n_threads);
+    fprintf(stderr, "  -p N,      --processors N      [%-7d] number of processors to use during computation\n", params.n_processors);
+    fprintf(stderr, "  -ot N,     --offset-t N        [%-7d] time offset in milliseconds\n",                    params.offset_t_ms);
+    fprintf(stderr, "  -on N,     --offset-n N        [%-7d] segment index offset\n",                           params.offset_n);
+    fprintf(stderr, "  -d  N,     --duration N        [%-7d] duration of audio to process in milliseconds\n",   params.duration_ms);
+    fprintf(stderr, "  -mc N,     --max-context N     [%-7d] maximum number of text context tokens to store\n", params.max_context);
+    fprintf(stderr, "  -ml N,     --max-len N         [%-7d] maximum segment length in characters\n",           params.max_len);
+    fprintf(stderr, "  -sow,      --split-on-word     [%-7s] split on word rather than on token\n",             params.split_on_word ? "true" : "false");
+    fprintf(stderr, "  -bo N,     --best-of N         [%-7d] number of best candidates to keep\n",              params.best_of);
+    fprintf(stderr, "  -bs N,     --beam-size N       [%-7d] beam size for beam search\n",                      params.beam_size);
+    fprintf(stderr, "  -wt N,     --word-thold N      [%-7.2f] word timestamp probability threshold\n",         params.word_thold);
+    fprintf(stderr, "  -et N,     --entropy-thold N   [%-7.2f] entropy threshold for decoder fail\n",           params.entropy_thold);
+    fprintf(stderr, "  -lpt N,    --logprob-thold N   [%-7.2f] log probability threshold for decoder fail\n",   params.logprob_thold);
+    fprintf(stderr, "  -su,       --speed-up          [%-7s] speed up audio by x2 (reduced accuracy)\n",        params.speed_up ? "true" : "false");
+    fprintf(stderr, "  -tr,       --translate         [%-7s] translate from source language to english\n",      params.translate ? "true" : "false");
+    fprintf(stderr, "  -di,       --diarize           [%-7s] stereo audio diarization\n",                       params.diarize ? "true" : "false");
+    fprintf(stderr, "  -nf,       --no-fallback       [%-7s] do not use temperature fallback while decoding\n", params.no_fallback ? "true" : "false");
+    fprintf(stderr, "  -otxt,     --output-txt        [%-7s] output result in a text file\n",                   params.output_txt ? "true" : "false");
+    fprintf(stderr, "  -ovtt,     --output-vtt        [%-7s] output result in a vtt file\n",                    params.output_vtt ? "true" : "false");
+    fprintf(stderr, "  -osrt,     --output-srt        [%-7s] output result in a srt file\n",                    params.output_srt ? "true" : "false");
+    fprintf(stderr, "  -olrc,     --output-lrc        [%-7s] output result in a lrc file\n",                    params.output_lrc ? "true" : "false");
+    fprintf(stderr, "  -owts,     --output-words      [%-7s] output script for generating karaoke video\n",     params.output_wts ? "true" : "false");
+    fprintf(stderr, "  -fp,       --font-path         [%-7s] path to a monospace font for karaoke video\n",     params.font_path.c_str());
+    fprintf(stderr, "  -ocsv,     --output-csv        [%-7s] output result in a CSV file\n",                    params.output_csv ? "true" : "false");
+    fprintf(stderr, "  -oj,       --output-json       [%-7s] output result in a JSON file\n",                   params.output_jsn ? "true" : "false");
+    fprintf(stderr, "  -of FNAME, --output-file FNAME [%-7s] output file path (without file extension)\n",      "");
+    fprintf(stderr, "  -ps,       --print-special     [%-7s] print special tokens\n",                           params.print_special ? "true" : "false");
+    fprintf(stderr, "  -pc,       --print-colors      [%-7s] print colors\n",                                   params.print_colors ? "true" : "false");
+    fprintf(stderr, "  -pp,       --print-progress    [%-7s] print progress\n",                                 params.print_progress ? "true" : "false");
+    fprintf(stderr, "  -nt,       --no-timestamps     [%-7s] do not print timestamps\n",                        params.no_timestamps ? "false" : "true");
+    fprintf(stderr, "  -l LANG,   --language LANG     [%-7s] spoken language ('auto' for auto-detect)\n",       params.language.c_str());
+    fprintf(stderr, "             --prompt PROMPT     [%-7s] initial prompt\n",                                 params.prompt.c_str());
+    fprintf(stderr, "  -m FNAME,  --model FNAME       [%-7s] model path\n",                                     params.model.c_str());
+    fprintf(stderr, "  -f FNAME,  --file FNAME        [%-7s] input WAV file path\n",                            "");
     fprintf(stderr, "\n");
 }
 
@@ -182,7 +203,7 @@ struct whisper_print_user_data {
     const std::vector<std::vector<float>> * pcmf32s;
 };
 
-void whisper_print_segment_callback(struct whisper_context * ctx, int n_new, void * user_data) {
+void whisper_print_segment_callback(struct whisper_context * ctx, struct whisper_state * /*state*/, int n_new, void * user_data) {
     const auto & params  = *((whisper_print_user_data *) user_data)->params;
     const auto & pcmf32s = *((whisper_print_user_data *) user_data)->pcmf32s;
 
@@ -190,8 +211,8 @@ void whisper_print_segment_callback(struct whisper_context * ctx, int n_new, voi
 
     std::string speaker = "";
 
-    int64_t t0;
-    int64_t t1;
+    int64_t t0 = 0;
+    int64_t t1 = 0;
 
     // print the last n_new segments
     const int s0 = n_segments - n_new;
@@ -341,31 +362,192 @@ bool output_csv(struct whisper_context * ctx, const char * fname) {
     fprintf(stderr, "%s: saving output to '%s'\n", __func__, fname);
 
     const int n_segments = whisper_full_n_segments(ctx);
+    fout << "start,end,text\n";
     for (int i = 0; i < n_segments; ++i) {
         const char * text = whisper_full_get_segment_text(ctx, i);
-        if (text[0] == ' ') {
-            text = text + sizeof(char); //whisper_full_get_segment_text() returns a string with leading space, point to the next character.
-        }
         const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
         const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
 
         //need to multiply times returned from whisper_full_get_segment_t{0,1}() by 10 to get milliseconds.
-        fout << 10 * t0 << ", " << 10 * t1 << ", \"" << text    << "\"\n";
+        fout << 10 * t0 << "," << 10 * t1 << ",\"" << text    << "\"\n";
     }
 
     return true;
 }
 
+char *escape_double_quotes(const char *str) {
+    if (str == NULL) {
+        return NULL;
+    }
+
+    size_t escaped_length = strlen(str) + 1;
+
+    for (size_t i = 0; str[i] != '\0'; i++) {
+        if (str[i] == '"') {
+            escaped_length++;
+        }
+    }
+
+    char *escaped = (char *)calloc(escaped_length, 1); // pre-zeroed
+    if (escaped == NULL) {
+        return NULL;
+    }
+
+    size_t pos = 0;
+    for (size_t i = 0; str[i] != '\0'; i++) {
+        if (str[i] == '"') {
+            escaped[pos++] = '\\';
+            escaped[pos++] = '"';
+        } else {
+            escaped[pos++] = str[i];
+        }
+    }
+
+    // no need to set zero due to calloc() being used prior
+
+    return escaped;
+}
+
+bool output_json(struct whisper_context * ctx, const char * fname, const whisper_params & params) {
+    std::ofstream fout(fname);
+    int indent = 0;
+
+    auto doindent = [&]() {
+        for (int i = 0; i < indent; i++) fout << "\t";
+    };
+
+    auto start_arr = [&](const char *name) {
+        doindent();
+        fout << "\"" << name << "\": [\n";
+        indent++;
+    };
+
+    auto end_arr = [&](bool end = false) {
+        indent--;
+        doindent();
+        fout << (end ? "]\n" : "},\n");
+    };
+
+    auto start_obj = [&](const char *name = nullptr) {
+        doindent();
+        if (name) {
+            fout << "\"" << name << "\": {\n";
+        } else {
+            fout << "{\n";
+        }
+        indent++;
+    };
+
+    auto end_obj = [&](bool end = false) {
+        indent--;
+        doindent();
+        fout << (end ? "}\n" : "},\n");
+    };
+
+    auto start_value = [&](const char *name) {
+        doindent();
+        fout << "\"" << name << "\": ";
+    };
+
+    auto value_s = [&](const char *name, const char *val, bool end = false) {
+        start_value(name);
+        char * val_escaped = escape_double_quotes(val);
+        fout << "\"" << val_escaped << (end ? "\"\n" : "\",\n");
+        free(val_escaped);
+    };
+
+    auto end_value = [&](bool end = false) {
+        fout << (end ? "\n" : ",\n");
+    };
+
+    auto value_i = [&](const char *name, const int64_t val, bool end = false) {
+        start_value(name);
+        fout << val;
+        end_value(end);
+    };
+
+    auto value_b = [&](const char *name, const bool val, bool end = false) {
+        start_value(name);
+        fout << (val ? "true" : "false");
+        end_value(end);
+    };
+
+    if (!fout.is_open()) {
+        fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname);
+        return false;
+    }
+
+    fprintf(stderr, "%s: saving output to '%s'\n", __func__, fname);
+    start_obj();
+        value_s("systeminfo", whisper_print_system_info());
+        start_obj("model");
+            value_s("type", whisper_model_type_readable(ctx));
+            value_b("multilingual", whisper_is_multilingual(ctx));
+            value_i("vocab", whisper_model_n_vocab(ctx));
+            start_obj("audio");
+                value_i("ctx", whisper_model_n_audio_ctx(ctx));
+                value_i("state", whisper_model_n_audio_state(ctx));
+                value_i("head", whisper_model_n_audio_head(ctx));
+                value_i("layer", whisper_model_n_audio_layer(ctx), true);
+            end_obj();
+            start_obj("text");
+                value_i("ctx", whisper_model_n_text_ctx(ctx));
+                value_i("state", whisper_model_n_text_state(ctx));
+                value_i("head", whisper_model_n_text_head(ctx));
+                value_i("layer", whisper_model_n_text_layer(ctx), true);
+            end_obj();
+            value_i("mels", whisper_model_n_mels(ctx));
+            value_i("f16", whisper_model_f16(ctx), true);
+        end_obj();
+        start_obj("params");
+            value_s("model", params.model.c_str());
+            value_s("language", params.language.c_str());
+            value_b("translate", params.translate, true);
+        end_obj();
+        start_obj("result");
+            value_s("language", whisper_lang_str(whisper_full_lang_id(ctx)), true);
+        end_obj();
+        start_arr("transcription");
+
+            const int n_segments = whisper_full_n_segments(ctx);
+            for (int i = 0; i < n_segments; ++i) {
+                const char * text = whisper_full_get_segment_text(ctx, i);
+                const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
+                const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
+
+                start_obj();
+                    start_obj("timestamps");
+                        value_s("from", to_timestamp(t0, true).c_str());
+                        value_s("to", to_timestamp(t1, true).c_str(), true);
+                    end_obj();
+                    start_obj("offsets");
+                        value_i("from", t0 * 10);
+                        value_i("to", t1 * 10, true);
+                    end_obj();
+                    value_s("text", text, true);
+                end_obj(i == (n_segments - 1));
+            }
+
+        end_arr(true);
+    end_obj(true);
+    return true;
+}
+
 // karaoke video generation
 // outputs a bash script that uses ffmpeg to generate a video with the subtitles
 // TODO: font parameter adjustments
-bool output_wts(struct whisper_context * ctx, const char * fname, const char * fname_inp, const whisper_params & /*params*/, float t_sec) {
+bool output_wts(struct whisper_context * ctx, const char * fname, const char * fname_inp, const whisper_params & params, float t_sec) {
     std::ofstream fout(fname);
 
     fprintf(stderr, "%s: saving output to '%s'\n", __func__, fname);
 
-    // TODO: become parameter
-    static const char * font = "/System/Library/Fonts/Supplemental/Courier New Bold.ttf";
+    static const char * font = params.font_path.c_str();
+
+    std::ifstream fin(font);
+    if (!fin.is_open()) {
+        fprintf(stderr, "%s: font not found at '%s', please specify a monospace font with -fp\n", __func__, font);
+        return false;
+    }
 
     fout << "#!/bin/bash" << "\n";
     fout << "\n";
@@ -468,6 +650,39 @@ bool output_wts(struct whisper_context * ctx, const char * fname, const char * f
     return true;
 }
 
+bool output_lrc(struct whisper_context * ctx, const char * fname) {
+
+    std::ofstream fout(fname);
+    if (!fout.is_open()) {
+        fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname);
+        return false;
+    }
+
+    fprintf(stderr, "%s: saving output to '%s'\n", __func__, fname);
+
+    fout << "[by:whisper.cpp]\n";
+
+    const int n_segments = whisper_full_n_segments(ctx);
+    for (int i = 0; i < n_segments; ++i) {
+        const char * text = whisper_full_get_segment_text(ctx, i);
+        const int64_t t = whisper_full_get_segment_t0(ctx, i);
+
+        int64_t msec = t * 10;
+        int64_t min = msec / (1000 * 60);
+        msec = msec - min * (1000 * 60);
+        int64_t sec = msec / 1000;
+        msec = msec - sec * 1000;
+
+        char buf[16];
+        snprintf(buf, sizeof(buf), "%02d:%02d.%02d", (int) min, (int) sec, (int) ( msec / 10));
+        std::string timestamp_lrc = std::string(buf);
+
+        fout <<  '[' << timestamp_lrc << ']' << text << "\n";
+    }
+
+    return true;
+}
+
 int main(int argc, char ** argv) {
     whisper_params params;
 
@@ -496,108 +711,16 @@ int main(int argc, char ** argv) {
         return 3;
     }
 
-    // initial prompt
-    std::vector<whisper_token> prompt_tokens;
-
-    if (!params.prompt.empty()) {
-        prompt_tokens.resize(1024);
-        prompt_tokens.resize(whisper_tokenize(ctx, params.prompt.c_str(), prompt_tokens.data(), prompt_tokens.size()));
-
-        fprintf(stderr, "\n");
-        fprintf(stderr, "initial prompt: '%s'\n", params.prompt.c_str());
-        fprintf(stderr, "initial tokens: [ ");
-        for (int i = 0; i < (int) prompt_tokens.size(); ++i) {
-            fprintf(stderr, "%d ", prompt_tokens[i]);
-        }
-        fprintf(stderr, "]\n");
-    }
-
     for (int f = 0; f < (int) params.fname_inp.size(); ++f) {
         const auto fname_inp = params.fname_inp[f];
+		const auto fname_out = f < (int) params.fname_out.size() && !params.fname_out[f].empty() ? params.fname_out[f] : params.fname_inp[f];
 
-        std::vector<float> pcmf32; // mono-channel F32 PCM
+        std::vector<float> pcmf32;               // mono-channel F32 PCM
         std::vector<std::vector<float>> pcmf32s; // stereo-channel F32 PCM
 
-        // WAV input
-        {
-            drwav wav;
-            std::vector<uint8_t> wav_data; // used for pipe input from stdin
-
-            if (fname_inp == "-") {
-                {
-                    uint8_t buf[1024];
-                    while (true)
-                    {
-                        const size_t n = fread(buf, 1, sizeof(buf), stdin);
-                        if (n == 0) {
-                            break;
-                        }
-                        wav_data.insert(wav_data.end(), buf, buf + n);
-                    }
-                }
-
-                if (drwav_init_memory(&wav, wav_data.data(), wav_data.size(), nullptr) == false) {
-                    fprintf(stderr, "error: failed to open WAV file from stdin\n");
-                    return 4;
-                }
-
-                fprintf(stderr, "%s: read %zu bytes from stdin\n", __func__, wav_data.size());
-            }
-            else if (drwav_init_file(&wav, fname_inp.c_str(), nullptr) == false) {
-                fprintf(stderr, "error: failed to open '%s' as WAV file\n", fname_inp.c_str());
-                return 5;
-            }
-
-            if (wav.channels != 1 && wav.channels != 2) {
-                fprintf(stderr, "%s: WAV file '%s' must be mono or stereo\n", argv[0], fname_inp.c_str());
-                return 6;
-            }
-
-            if (params.diarize && wav.channels != 2 && params.no_timestamps == false) {
-                fprintf(stderr, "%s: WAV file '%s' must be stereo for diarization and timestamps have to be enabled\n", argv[0], fname_inp.c_str());
-                return 6;
-            }
-
-            if (wav.sampleRate != WHISPER_SAMPLE_RATE) {
-                fprintf(stderr, "%s: WAV file '%s' must be %i kHz\n", argv[0], fname_inp.c_str(), WHISPER_SAMPLE_RATE/1000);
-                return 8;
-            }
-
-            if (wav.bitsPerSample != 16) {
-                fprintf(stderr, "%s: WAV file '%s' must be 16-bit\n", argv[0], fname_inp.c_str());
-                return 9;
-            }
-
-            const uint64_t n = wav_data.empty() ? wav.totalPCMFrameCount : wav_data.size()/(wav.channels*wav.bitsPerSample/8);
-
-            std::vector<int16_t> pcm16;
-            pcm16.resize(n*wav.channels);
-            drwav_read_pcm_frames_s16(&wav, n, pcm16.data());
-            drwav_uninit(&wav);
-
-            // convert to mono, float
-            pcmf32.resize(n);
-            if (wav.channels == 1) {
-                for (uint64_t i = 0; i < n; i++) {
-                    pcmf32[i] = float(pcm16[i])/32768.0f;
-                }
-            } else {
-                for (uint64_t i = 0; i < n; i++) {
-                    pcmf32[i] = float(pcm16[2*i] + pcm16[2*i + 1])/65536.0f;
-                }
-            }
-
-            if (params.diarize) {
-                // convert to stereo, float
-                pcmf32s.resize(2);
-
-                pcmf32s[0].resize(n);
-                pcmf32s[1].resize(n);
-                for (uint64_t i = 0; i < n; i++) {
-                    pcmf32s[0][i] = float(pcm16[2*i])/32768.0f;
-                    pcmf32s[1][i] = float(pcm16[2*i + 1])/32768.0f;
-                }
-            }
+        if (!::read_wav(fname_inp, pcmf32, pcmf32s, params.diarize)) {
+            fprintf(stderr, "error: failed to read WAV file '%s'\n", fname_inp.c_str());
+            continue;
         }
 
         // print system information
@@ -646,18 +769,19 @@ int main(int argc, char ** argv) {
 
             wparams.token_timestamps = params.output_wts || params.max_len > 0;
             wparams.thold_pt         = params.word_thold;
-            wparams.entropy_thold    = params.entropy_thold;
-            wparams.logprob_thold    = params.logprob_thold;
             wparams.max_len          = params.output_wts && params.max_len == 0 ? 60 : params.max_len;
+            wparams.split_on_word    = params.split_on_word;
 
             wparams.speed_up         = params.speed_up;
 
+            wparams.initial_prompt   = params.prompt.c_str();
+
             wparams.greedy.best_of        = params.best_of;
             wparams.beam_search.beam_size = params.beam_size;
-            wparams.temperature_inc = -1;
 
-            wparams.prompt_tokens     = prompt_tokens.empty() ? nullptr : prompt_tokens.data();
-            wparams.prompt_n_tokens   = prompt_tokens.empty() ? 0       : prompt_tokens.size();
+            wparams.temperature_inc  = params.no_fallback ? 0.0f : wparams.temperature_inc;
+            wparams.entropy_thold    = params.entropy_thold;
+            wparams.logprob_thold    = params.logprob_thold;
 
             whisper_print_user_data user_data = { &params, &pcmf32s };
 
@@ -673,7 +797,7 @@ int main(int argc, char ** argv) {
             {
                 static bool is_aborted = false; // NOTE: this should be atomic to avoid data race
 
-                wparams.encoder_begin_callback = [](struct whisper_context * /*ctx*/, void * user_data) {
+                wparams.encoder_begin_callback = [](struct whisper_context * /*ctx*/, struct whisper_state * /*state*/, void * user_data) {
                     bool is_aborted = *(bool*)user_data;
                     return !is_aborted;
                 };
@@ -692,34 +816,45 @@ int main(int argc, char ** argv) {
 
             // output to text file
             if (params.output_txt) {
-                const auto fname_txt = fname_inp + ".txt";
+                const auto fname_txt = fname_out + ".txt";
                 output_txt(ctx, fname_txt.c_str());
             }
 
             // output to VTT file
             if (params.output_vtt) {
-                const auto fname_vtt = fname_inp + ".vtt";
+                const auto fname_vtt = fname_out + ".vtt";
                 output_vtt(ctx, fname_vtt.c_str());
             }
 
             // output to SRT file
             if (params.output_srt) {
-                const auto fname_srt = fname_inp + ".srt";
+                const auto fname_srt = fname_out + ".srt";
                 output_srt(ctx, fname_srt.c_str(), params);
             }
 
             // output to WTS file
             if (params.output_wts) {
-                const auto fname_wts = fname_inp + ".wts";
+                const auto fname_wts = fname_out + ".wts";
                 output_wts(ctx, fname_wts.c_str(), fname_inp.c_str(), params, float(pcmf32.size() + 1000)/WHISPER_SAMPLE_RATE);
             }
 
-	    // output to CSV file
+            // output to CSV file
             if (params.output_csv) {
-                const auto fname_csv = fname_inp + ".csv";
+                const auto fname_csv = fname_out + ".csv";
                 output_csv(ctx, fname_csv.c_str());
             }
 
+            // output to JSON file
+            if (params.output_jsn) {
+                const auto fname_jsn = fname_out + ".json";
+                output_json(ctx, fname_jsn.c_str(), params);
+            }
+
+            // output to LRC file
+            if (params.output_lrc) {
+                const auto fname_lrc = fname_out + ".lrc";
+                output_lrc(ctx, fname_lrc.c_str());
+            }
         }
     }
 

examples/stream/CMakeLists.txt

@@ -5,6 +5,5 @@ if (WHISPER_SUPPORT_SDL2)
 
     include(DefaultTargetOptions)
 
-    target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS})
-    target_link_libraries(${TARGET} PRIVATE whisper ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
+    target_link_libraries(${TARGET} PRIVATE common common-sdl whisper ${CMAKE_THREAD_LIBS_INIT})
 endif ()

examples/stream/stream.cpp

@@ -3,19 +3,16 @@
 // A very quick-n-dirty implementation serving mainly as a proof of concept.
 //
 
+#include "common.h"
+#include "common-sdl.h"
 #include "whisper.h"
 
-#include <SDL.h>
-#include <SDL_audio.h>
-
-#include <atomic>
 #include <cassert>
 #include <cstdio>
 #include <string>
 #include <thread>
 #include <vector>
 #include <fstream>
-#include <mutex>
 
 //  500 -> 00:05.000
 // 6000 -> 01:00.000
@@ -46,6 +43,7 @@ struct whisper_params {
 
     bool speed_up      = false;
     bool translate     = false;
+    bool no_fallback   = false;
     bool print_special = false;
     bool no_context    = true;
     bool no_timestamps = false;
@@ -76,6 +74,7 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
         else if (arg == "-fth" || arg == "--freq-thold")    { params.freq_thold    = std::stof(argv[++i]); }
         else if (arg == "-su"  || arg == "--speed-up")      { params.speed_up      = true; }
         else if (arg == "-tr"  || arg == "--translate")     { params.translate     = true; }
+        else if (arg == "-nf"  || arg == "--no-fallback")   { params.no_fallback   = true; }
         else if (arg == "-ps"  || arg == "--print-special") { params.print_special = true; }
         else if (arg == "-kc"  || arg == "--keep-context")  { params.no_context    = false; }
         else if (arg == "-l"   || arg == "--language")      { params.language      = argv[++i]; }
@@ -97,325 +96,26 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "\n");
     fprintf(stderr, "options:\n");
     fprintf(stderr, "  -h,       --help          [default] show this help message and exit\n");
-    fprintf(stderr, "  -t N,     --threads N     [%-7d] number of threads to use during computation\n", params.n_threads);
-    fprintf(stderr, "            --step N        [%-7d] audio step size in milliseconds\n",             params.step_ms);
-    fprintf(stderr, "            --length N      [%-7d] audio length in milliseconds\n",                params.length_ms);
-    fprintf(stderr, "            --keep N        [%-7d] audio to keep from previous step in ms\n",      params.keep_ms);
-    fprintf(stderr, "  -c ID,    --capture ID    [%-7d] capture device ID\n",                           params.capture_id);
-    fprintf(stderr, "  -mt N,    --max-tokens N  [%-7d] maximum number of tokens per audio chunk\n",    params.max_tokens);
-    fprintf(stderr, "  -ac N,    --audio-ctx N   [%-7d] audio context size (0 - all)\n",                params.audio_ctx);
-    fprintf(stderr, "  -vth N,   --vad-thold N   [%-7.2f] voice activity detection threshold\n",        params.vad_thold);
-    fprintf(stderr, "  -fth N,   --freq-thold N  [%-7.2f] high-pass frequency cutoff\n",                params.freq_thold);
-    fprintf(stderr, "  -su,      --speed-up      [%-7s] speed up audio by x2 (reduced accuracy)\n",     params.speed_up ? "true" : "false");
-    fprintf(stderr, "  -tr,      --translate     [%-7s] translate from source language to english\n",   params.translate ? "true" : "false");
-    fprintf(stderr, "  -ps,      --print-special [%-7s] print special tokens\n",                        params.print_special ? "true" : "false");
-    fprintf(stderr, "  -kc,      --keep-context  [%-7s] keep context between audio chunks\n",           params.no_context ? "false" : "true");
-    fprintf(stderr, "  -l LANG,  --language LANG [%-7s] spoken language\n",                             params.language.c_str());
-    fprintf(stderr, "  -m FNAME, --model FNAME   [%-7s] model path\n",                                  params.model.c_str());
-    fprintf(stderr, "  -f FNAME, --file FNAME    [%-7s] text output file name\n",                       params.fname_out.c_str());
+    fprintf(stderr, "  -t N,     --threads N     [%-7d] number of threads to use during computation\n",    params.n_threads);
+    fprintf(stderr, "            --step N        [%-7d] audio step size in milliseconds\n",                params.step_ms);
+    fprintf(stderr, "            --length N      [%-7d] audio length in milliseconds\n",                   params.length_ms);
+    fprintf(stderr, "            --keep N        [%-7d] audio to keep from previous step in ms\n",         params.keep_ms);
+    fprintf(stderr, "  -c ID,    --capture ID    [%-7d] capture device ID\n",                              params.capture_id);
+    fprintf(stderr, "  -mt N,    --max-tokens N  [%-7d] maximum number of tokens per audio chunk\n",       params.max_tokens);
+    fprintf(stderr, "  -ac N,    --audio-ctx N   [%-7d] audio context size (0 - all)\n",                   params.audio_ctx);
+    fprintf(stderr, "  -vth N,   --vad-thold N   [%-7.2f] voice activity detection threshold\n",           params.vad_thold);
+    fprintf(stderr, "  -fth N,   --freq-thold N  [%-7.2f] high-pass frequency cutoff\n",                   params.freq_thold);
+    fprintf(stderr, "  -su,      --speed-up      [%-7s] speed up audio by x2 (reduced accuracy)\n",        params.speed_up ? "true" : "false");
+    fprintf(stderr, "  -tr,      --translate     [%-7s] translate from source language to english\n",      params.translate ? "true" : "false");
+    fprintf(stderr, "  -nf,      --no-fallback   [%-7s] do not use temperature fallback while decoding\n", params.no_fallback ? "true" : "false");
+    fprintf(stderr, "  -ps,      --print-special [%-7s] print special tokens\n",                           params.print_special ? "true" : "false");
+    fprintf(stderr, "  -kc,      --keep-context  [%-7s] keep context between audio chunks\n",              params.no_context ? "false" : "true");
+    fprintf(stderr, "  -l LANG,  --language LANG [%-7s] spoken language\n",                                params.language.c_str());
+    fprintf(stderr, "  -m FNAME, --model FNAME   [%-7s] model path\n",                                     params.model.c_str());
+    fprintf(stderr, "  -f FNAME, --file FNAME    [%-7s] text output file name\n",                          params.fname_out.c_str());
     fprintf(stderr, "\n");
 }
 
-//
-// SDL Audio capture
-//
-
-class audio_async {
-public:
-    audio_async(int len_ms);
-    ~audio_async();
-
-    bool init(int capture_id, int sample_rate);
-
-    // start capturing audio via the provided SDL callback
-    // keep last len_ms seconds of audio in a circular buffer
-    bool resume();
-    bool pause();
-    bool clear();
-
-    // callback to be called by SDL
-    void callback(uint8_t * stream, int len);
-
-    // get audio data from the circular buffer
-    void get(int ms, std::vector<float> & audio);
-
-private:
-    SDL_AudioDeviceID m_dev_id_in = 0;
-
-    int m_len_ms = 0;
-    int m_sample_rate = 0;
-
-    std::atomic_bool m_running;
-    std::mutex       m_mutex;
-
-    std::vector<float> m_audio;
-    std::vector<float> m_audio_new;
-    size_t             m_audio_pos = 0;
-    size_t             m_audio_len = 0;
-};
-
-audio_async::audio_async(int len_ms) {
-    m_len_ms = len_ms;
-
-    m_running = false;
-}
-
-audio_async::~audio_async() {
-    if (m_dev_id_in) {
-        SDL_CloseAudioDevice(m_dev_id_in);
-    }
-}
-
-bool audio_async::init(int capture_id, int sample_rate) {
-    SDL_LogSetPriority(SDL_LOG_CATEGORY_APPLICATION, SDL_LOG_PRIORITY_INFO);
-
-    if (SDL_Init(SDL_INIT_AUDIO) < 0) {
-        SDL_LogError(SDL_LOG_CATEGORY_APPLICATION, "Couldn't initialize SDL: %s\n", SDL_GetError());
-        return false;
-    }
-
-    SDL_SetHintWithPriority(SDL_HINT_AUDIO_RESAMPLING_MODE, "medium", SDL_HINT_OVERRIDE);
-
-    {
-        int nDevices = SDL_GetNumAudioDevices(SDL_TRUE);
-        fprintf(stderr, "%s: found %d capture devices:\n", __func__, nDevices);
-        for (int i = 0; i < nDevices; i++) {
-            fprintf(stderr, "%s:    - Capture device #%d: '%s'\n", __func__, i, SDL_GetAudioDeviceName(i, SDL_TRUE));
-        }
-    }
-
-    SDL_AudioSpec capture_spec_requested;
-    SDL_AudioSpec capture_spec_obtained;
-
-    SDL_zero(capture_spec_requested);
-    SDL_zero(capture_spec_obtained);
-
-    capture_spec_requested.freq     = sample_rate;
-    capture_spec_requested.format   = AUDIO_F32;
-    capture_spec_requested.channels = 1;
-    capture_spec_requested.samples  = 1024;
-    capture_spec_requested.callback = [](void * userdata, uint8_t * stream, int len) {
-        audio_async * audio = (audio_async *) userdata;
-        audio->callback(stream, len);
-    };
-    capture_spec_requested.userdata = this;
-
-    if (capture_id >= 0) {
-        fprintf(stderr, "%s: attempt to open capture device %d : '%s' ...\n", __func__, capture_id, SDL_GetAudioDeviceName(capture_id, SDL_TRUE));
-        m_dev_id_in = SDL_OpenAudioDevice(SDL_GetAudioDeviceName(capture_id, SDL_TRUE), SDL_TRUE, &capture_spec_requested, &capture_spec_obtained, 0);
-    } else {
-        fprintf(stderr, "%s: attempt to open default capture device ...\n", __func__);
-        m_dev_id_in = SDL_OpenAudioDevice(nullptr, SDL_TRUE, &capture_spec_requested, &capture_spec_obtained, 0);
-    }
-
-    if (!m_dev_id_in) {
-        fprintf(stderr, "%s: couldn't open an audio device for capture: %s!\n", __func__, SDL_GetError());
-        m_dev_id_in = 0;
-
-        return false;
-    } else {
-        fprintf(stderr, "%s: obtained spec for input device (SDL Id = %d):\n", __func__, m_dev_id_in);
-        fprintf(stderr, "%s:     - sample rate:       %d\n",                   __func__, capture_spec_obtained.freq);
-        fprintf(stderr, "%s:     - format:            %d (required: %d)\n",    __func__, capture_spec_obtained.format,
-                capture_spec_requested.format);
-        fprintf(stderr, "%s:     - channels:          %d (required: %d)\n",    __func__, capture_spec_obtained.channels,
-                capture_spec_requested.channels);
-        fprintf(stderr, "%s:     - samples per frame: %d\n",                   __func__, capture_spec_obtained.samples);
-    }
-
-    m_sample_rate = capture_spec_obtained.freq;
-
-    m_audio.resize((m_sample_rate*m_len_ms)/1000);
-
-    return true;
-}
-
-bool audio_async::resume() {
-    if (!m_dev_id_in) {
-        fprintf(stderr, "%s: no audio device to resume!\n", __func__);
-        return false;
-    }
-
-    if (m_running) {
-        fprintf(stderr, "%s: already running!\n", __func__);
-        return false;
-    }
-
-    SDL_PauseAudioDevice(m_dev_id_in, 0);
-
-    m_running = true;
-
-    return true;
-}
-
-bool audio_async::pause() {
-    if (!m_dev_id_in) {
-        fprintf(stderr, "%s: no audio device to pause!\n", __func__);
-        return false;
-    }
-
-    if (!m_running) {
-        fprintf(stderr, "%s: already paused!\n", __func__);
-        return false;
-    }
-
-    SDL_PauseAudioDevice(m_dev_id_in, 1);
-
-    m_running = false;
-
-    return true;
-}
-
-bool audio_async::clear() {
-    if (!m_dev_id_in) {
-        fprintf(stderr, "%s: no audio device to clear!\n", __func__);
-        return false;
-    }
-
-    if (!m_running) {
-        fprintf(stderr, "%s: not running!\n", __func__);
-        return false;
-    }
-
-    {
-        std::lock_guard<std::mutex> lock(m_mutex);
-
-        m_audio_pos = 0;
-        m_audio_len = 0;
-    }
-
-    return true;
-}
-
-// callback to be called by SDL
-void audio_async::callback(uint8_t * stream, int len) {
-    if (!m_running) {
-        return;
-    }
-
-    const size_t n_samples = len / sizeof(float);
-
-    m_audio_new.resize(n_samples);
-    memcpy(m_audio_new.data(), stream, n_samples * sizeof(float));
-
-    //fprintf(stderr, "%s: %zu samples, pos %zu, len %zu\n", __func__, n_samples, m_audio_pos, m_audio_len);
-
-    {
-        std::lock_guard<std::mutex> lock(m_mutex);
-
-        if (m_audio_pos + n_samples > m_audio.size()) {
-            const size_t n0 = m_audio.size() - m_audio_pos;
-
-            memcpy(&m_audio[m_audio_pos], stream, n0 * sizeof(float));
-            memcpy(&m_audio[0], &stream[n0], (n_samples - n0) * sizeof(float));
-
-            m_audio_pos = (m_audio_pos + n_samples) % m_audio.size();
-            m_audio_len = m_audio.size();
-        } else {
-            memcpy(&m_audio[m_audio_pos], stream, n_samples * sizeof(float));
-
-            m_audio_pos = (m_audio_pos + n_samples) % m_audio.size();
-            m_audio_len = std::min(m_audio_len + n_samples, m_audio.size());
-        }
-    }
-}
-
-void audio_async::get(int ms, std::vector<float> & result) {
-    if (!m_dev_id_in) {
-        fprintf(stderr, "%s: no audio device to get audio from!\n", __func__);
-        return;
-    }
-
-    if (!m_running) {
-        fprintf(stderr, "%s: not running!\n", __func__);
-        return;
-    }
-
-    result.clear();
-
-    {
-        std::lock_guard<std::mutex> lock(m_mutex);
-
-        if (ms <= 0) {
-            ms = m_len_ms;
-        }
-
-        size_t n_samples = (m_sample_rate * ms) / 1000;
-        if (n_samples > m_audio_len) {
-            n_samples = m_audio_len;
-        }
-
-        result.resize(n_samples);
-
-        int s0 = m_audio_pos - n_samples;
-        if (s0 < 0) {
-            s0 += m_audio.size();
-        }
-
-        if (s0 + n_samples > m_audio.size()) {
-            const size_t n0 = m_audio.size() - s0;
-
-            memcpy(result.data(), &m_audio[s0], n0 * sizeof(float));
-            memcpy(&result[n0], &m_audio[0], (n_samples - n0) * sizeof(float));
-        } else {
-            memcpy(result.data(), &m_audio[s0], n_samples * sizeof(float));
-        }
-    }
-}
-
-///////////////////////////
-
-void high_pass_filter(std::vector<float> & data, float cutoff, float sample_rate) {
-    const float rc = 1.0f / (2.0f * M_PI * cutoff);
-    const float dt = 1.0f / sample_rate;
-    const float alpha = dt / (rc + dt);
-
-    float y = data[0];
-
-    for (size_t i = 1; i < data.size(); i++) {
-        y = alpha * (y + data[i] - data[i - 1]);
-        data[i] = y;
-    }
-}
-
-bool vad_simple(std::vector<float> & pcmf32, int sample_rate, int last_ms, float vad_thold, float freq_thold, bool verbose) {
-    const int n_samples      = pcmf32.size();
-    const int n_samples_last = (sample_rate * last_ms) / 1000;
-
-    if (n_samples_last >= n_samples) {
-        // not enough samples - assume no speech
-        return false;
-    }
-
-    if (freq_thold > 0.0f) {
-        high_pass_filter(pcmf32, freq_thold, sample_rate);
-    }
-
-    float energy_all  = 0.0f;
-    float energy_last = 0.0f;
-
-    for (int i = 0; i < n_samples; i++) {
-        energy_all += fabsf(pcmf32[i]);
-
-        if (i >= n_samples - n_samples_last) {
-            energy_last += fabsf(pcmf32[i]);
-        }
-    }
-
-    energy_all  /= n_samples;
-    energy_last /= n_samples_last;
-
-    if (verbose) {
-        fprintf(stderr, "%s: energy_all: %f, energy_last: %f, vad_thold: %f, freq_thold: %f\n", __func__, energy_all, energy_last, vad_thold, freq_thold);
-    }
-
-    if (energy_last > vad_thold*energy_all) {
-        return false;
-    }
-
-    return true;
-}
-
 int main(int argc, char ** argv) {
     whisper_params params;
 
@@ -423,20 +123,21 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    params.keep_ms = std::min(params.keep_ms, params.step_ms); // cannot be more than step_ms
+    params.keep_ms   = std::min(params.keep_ms,   params.step_ms);
+    params.length_ms = std::max(params.length_ms, params.step_ms);
 
-    const int n_samples_step = (params.step_ms  *1e-3)*WHISPER_SAMPLE_RATE;
-    const int n_samples_len  = (params.length_ms*1e-3)*WHISPER_SAMPLE_RATE;
-    const int n_samples_keep = (params.keep_ms  *1e-3)*WHISPER_SAMPLE_RATE;
-    const int n_samples_30s  = (30000           *1e-3)*WHISPER_SAMPLE_RATE;
+    const int n_samples_step = (1e-3*params.step_ms  )*WHISPER_SAMPLE_RATE;
+    const int n_samples_len  = (1e-3*params.length_ms)*WHISPER_SAMPLE_RATE;
+    const int n_samples_keep = (1e-3*params.keep_ms  )*WHISPER_SAMPLE_RATE;
+    const int n_samples_30s  = (1e-3*30000.0         )*WHISPER_SAMPLE_RATE;
 
     const bool use_vad = n_samples_step <= 0; // sliding window mode uses VAD
 
-    const int n_new_line = !use_vad ? params.length_ms / params.step_ms - 1 : 1; // number of steps to print new line
+    const int n_new_line = !use_vad ? std::max(1, params.length_ms / params.step_ms - 1) : 1; // number of steps to print new line
 
-    params.no_timestamps = !use_vad;
-    params.no_context    = use_vad;
-    params.max_tokens    = 0;
+    params.no_timestamps  = !use_vad;
+    params.no_context    |= use_vad;
+    params.max_tokens     = 0;
 
     // init audio
 
@@ -450,7 +151,7 @@ int main(int argc, char ** argv) {
 
     // whisper init
 
-    if (whisper_lang_id(params.language.c_str()) == -1) {
+    if (params.language != "auto" && whisper_lang_id(params.language.c_str()) == -1){
         fprintf(stderr, "error: unknown language '%s'\n", params.language.c_str());
         whisper_print_usage(argc, argv, params);
         exit(0);
@@ -486,7 +187,7 @@ int main(int argc, char ** argv) {
                 params.no_timestamps ? 0 : 1);
 
         if (!use_vad) {
-            fprintf(stderr, "%s: n_new_line = %d\n", __func__, n_new_line);
+            fprintf(stderr, "%s: n_new_line = %d, no_context = %d\n", __func__, n_new_line, params.no_context);
         } else {
             fprintf(stderr, "%s: using VAD, will transcribe on speech activity\n", __func__);
         }
@@ -516,23 +217,7 @@ int main(int argc, char ** argv) {
     // main audio loop
     while (is_running) {
         // handle Ctrl + C
-        {
-            SDL_Event event;
-            while (SDL_PollEvent(&event)) {
-                switch (event.type) {
-                    case SDL_QUIT:
-                        {
-                            is_running = false;
-                        } break;
-                    default:
-                        break;
-                }
-            }
-
-            if (!is_running) {
-                break;
-            }
-        }
+        is_running = sdl_poll_events();
 
         if (!is_running) {
             break;
@@ -555,7 +240,7 @@ int main(int argc, char ** argv) {
                     break;
                 }
 
-                SDL_Delay(1);
+                std::this_thread::sleep_for(std::chrono::milliseconds(1));
             }
 
             const int n_samples_new = pcmf32_new.size();
@@ -586,7 +271,7 @@ int main(int argc, char ** argv) {
 
             audio.get(2000, pcmf32_new);
 
-            if (vad_simple(pcmf32_new, WHISPER_SAMPLE_RATE, 1000, params.vad_thold, params.freq_thold, false)) {
+            if (::vad_simple(pcmf32_new, WHISPER_SAMPLE_RATE, 1000, params.vad_thold, params.freq_thold, false)) {
                 audio.get(params.length_ms, pcmf32);
             } else {
                 std::this_thread::sleep_for(std::chrono::milliseconds(100));
@@ -606,7 +291,6 @@ int main(int argc, char ** argv) {
             wparams.print_realtime   = false;
             wparams.print_timestamps = !params.no_timestamps;
             wparams.translate        = params.translate;
-            wparams.no_context       = true;
             wparams.single_segment   = !use_vad;
             wparams.max_tokens       = params.max_tokens;
             wparams.language         = params.language.c_str();
@@ -616,7 +300,8 @@ int main(int argc, char ** argv) {
             wparams.speed_up         = params.speed_up;
 
             // disable temperature fallback
-            wparams.temperature_inc  = -1.0f;
+            //wparams.temperature_inc  = -1.0f;
+            wparams.temperature_inc  = params.no_fallback ? 0.0f : wparams.temperature_inc;
 
             wparams.prompt_tokens    = params.no_context ? nullptr : prompt_tokens.data();
             wparams.prompt_n_tokens  = params.no_context ? 0       : prompt_tokens.size();

examples/talk-llama/.gitignore

@@ -0,0 +1 @@
+audio.mp3

examples/talk-llama/CMakeLists.txt

@@ -0,0 +1,16 @@
+if (WHISPER_SUPPORT_SDL2)
+    # talk-llama
+    set(TARGET talk-llama)
+    #add_executable(${TARGET} talk-llama.cpp llama.cpp)
+    #target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS})
+    #target_link_libraries(${TARGET} PRIVATE common common-sdl whisper ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
+
+    # TODO: this is temporary
+    #       need to export ggml symbols for MSVC, but too lazy ..
+    add_executable(${TARGET} talk-llama.cpp llama.cpp ../common.cpp ../common-sdl.cpp ../../ggml.c ../../whisper.cpp)
+
+    target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS} ../../)
+    target_link_libraries(${TARGET} PRIVATE ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
+
+    include(DefaultTargetOptions)
+endif ()

examples/talk-llama/README.md

@@ -0,0 +1,36 @@
+# talk-llama
+
+Talk with an LLaMA AI in your terminal
+
+[Demo Talk](https://user-images.githubusercontent.com/1991296/228024237-848f998c-c334-46a6-bef8-3271590da83b.mp4)
+
+## Building
+
+The `talk-llama` tool depends on SDL2 library to capture audio from the microphone. You can build it like this:
+
+```bash
+# Install SDL2 on Linux
+sudo apt-get install libsdl2-dev
+
+# Install SDL2 on Mac OS
+brew install sdl2
+
+# Build the "talk-llama" executable
+make talk-llama
+
+# Run it
+./talk-llama -mw ./models/ggml-small.en.bin -ml ../llama.cpp/models/13B/ggml-model-q4_0.bin -p "Georgi" -t 8
+```
+
+- The `-mw` argument specifies the Whisper model that you would like to use. Recommended `base` or `small` for real-time experience
+- The `-ml` argument specifies the LLaMA model that you would like to use. Read the instructions in https://github.com/ggerganov/llama.cpp for information about how to obtain a `ggml` compatible LLaMA model
+
+## TTS
+
+For best experience, this example needs a TTS tool to convert the generated text responses to voice.
+You can use any TTS engine that you would like - simply edit the [speak.sh](speak.sh) script to your needs.
+By default, it is configured to use MacOS's `say`, but you can use whatever you wish.
+
+## Discussion
+
+If you have any feedback, please let "us" know in the following discussion: https://github.com/ggerganov/whisper.cpp/discussions/672?converting=1

examples/talk-llama/eleven-labs.py

@@ -0,0 +1,23 @@
+import sys
+import importlib.util
+
+api_key = "" #Write your https://beta.elevenlabs.io api key here
+if not api_key:
+    print("To use elevenlabs you have to register to https://beta.elevenlabs.io and add your elevenlabs api key to examples/talk-llama/eleven-labs.py")
+    sys.exit()
+
+if importlib.util.find_spec("elevenlabs") is None:
+    print("elevenlabs library is not installed, you can install it to your enviroment using 'pip install elevenlabs'")
+    sys.exit()
+
+from elevenlabs import ElevenLabs
+eleven = ElevenLabs(api_key)
+
+# Get a Voice object, by name or UUID
+voice = eleven.voices["Arnold"] #Possible Voices: Adam Antoni Arnold Bella Domi Elli Josh
+
+# Generate the TTS
+audio = voice.generate(str(sys.argv[2:]))
+
+# Save the TTS to a file
+audio.save("audio") 

examples/talk-llama/llama.h

@@ -0,0 +1,173 @@
+#ifndef LLAMA_H
+#define LLAMA_H
+
+#include <stddef.h>
+#include <stdint.h>
+#include <stdbool.h>
+
+#ifdef LLAMA_SHARED
+#    if defined(_WIN32) && !defined(__MINGW32__)
+#        ifdef LLAMA_BUILD
+#            define LLAMA_API __declspec(dllexport)
+#        else
+#            define LLAMA_API __declspec(dllimport)
+#        endif
+#    else
+#        define LLAMA_API __attribute__ ((visibility ("default")))
+#    endif
+#else
+#    define LLAMA_API
+#endif
+
+#define LLAMA_FILE_VERSION 1
+#define LLAMA_FILE_MAGIC 0x67676a74 // 'ggjt' in hex
+#define LLAMA_FILE_MAGIC_UNVERSIONED 0x67676d6c // pre-versioned files
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+    //
+    // C interface
+    //
+    // TODO: show sample usage
+    //
+
+    struct llama_context;
+
+    typedef int llama_token;
+
+    typedef struct llama_token_data {
+        llama_token id;  // token id
+
+        float p;     // probability of the token
+        float plog;  // log probability of the token
+
+    } llama_token_data;
+
+    typedef void (*llama_progress_callback)(float progress, void *ctx);
+
+    struct llama_context_params {
+        int n_ctx;   // text context
+        int n_parts; // -1 for default
+        int seed;    // RNG seed, 0 for random
+
+        bool f16_kv;     // use fp16 for KV cache
+        bool logits_all; // the llama_eval() call computes all logits, not just the last one
+        bool vocab_only; // only load the vocabulary, no weights
+        bool use_mmap;   // use mmap if possible
+        bool use_mlock;  // force system to keep model in RAM
+        bool embedding;  // embedding mode only
+
+        // called with a progress value between 0 and 1, pass NULL to disable
+        llama_progress_callback progress_callback;
+        // context pointer passed to the progress callback
+        void * progress_callback_user_data;
+    };
+
+    LLAMA_API struct llama_context_params llama_context_default_params();
+
+    LLAMA_API bool llama_mmap_supported();
+    LLAMA_API bool llama_mlock_supported();
+
+    // Various functions for loading a ggml llama model.
+    // Allocate (almost) all memory needed for the model.
+    // Return NULL on failure
+    LLAMA_API struct llama_context * llama_init_from_file(
+                             const char * path_model,
+            struct llama_context_params   params);
+
+    // Frees all allocated memory
+    LLAMA_API void llama_free(struct llama_context * ctx);
+
+    // TODO: not great API - very likely to change
+    // Returns 0 on success
+    LLAMA_API int llama_model_quantize(
+            const char * fname_inp,
+            const char * fname_out,
+                   int   itype);
+
+    // Returns the KV cache that will contain the context for the
+    // ongoing prediction with the model.
+    LLAMA_API const uint8_t * llama_get_kv_cache(struct llama_context * ctx);
+
+    // Returns the size of the KV cache
+    LLAMA_API size_t llama_get_kv_cache_size(struct llama_context * ctx);
+
+    // Returns the number of tokens in the KV cache
+    LLAMA_API int llama_get_kv_cache_token_count(struct llama_context * ctx);
+
+    // Sets the KV cache containing the current context for the model
+    LLAMA_API void llama_set_kv_cache(
+            struct llama_context * ctx,
+                   const uint8_t * kv_cache,
+                          size_t   n_size,
+                             int   n_token_count);
+
+    // Run the llama inference to obtain the logits and probabilities for the next token.
+    // tokens + n_tokens is the provided batch of new tokens to process
+    // n_past is the number of tokens to use from previous eval calls
+    // Returns 0 on success
+    LLAMA_API int llama_eval(
+            struct llama_context * ctx,
+               const llama_token * tokens,
+                             int   n_tokens,
+                             int   n_past,
+                             int   n_threads);
+
+    // Convert the provided text into tokens.
+    // The tokens pointer must be large enough to hold the resulting tokens.
+    // Returns the number of tokens on success, no more than n_max_tokens
+    // Returns a negative number on failure - the number of tokens that would have been returned
+    // TODO: not sure if correct
+    LLAMA_API int llama_tokenize(
+            struct llama_context * ctx,
+                      const char * text,
+                     llama_token * tokens,
+                             int   n_max_tokens,
+                            bool   add_bos);
+
+    LLAMA_API int llama_n_vocab(struct llama_context * ctx);
+    LLAMA_API int llama_n_ctx  (struct llama_context * ctx);
+    LLAMA_API int llama_n_embd (struct llama_context * ctx);
+
+    // Token logits obtained from the last call to llama_eval()
+    // The logits for the last token are stored in the last row
+    // Can be mutated in order to change the probabilities of the next token
+    // Rows: n_tokens
+    // Cols: n_vocab
+    LLAMA_API float * llama_get_logits(struct llama_context * ctx);
+
+    // Get the embeddings for the input
+    // shape: [n_embd] (1-dimensional)
+    LLAMA_API float * llama_get_embeddings(struct llama_context * ctx);
+
+    // Token Id -> String. Uses the vocabulary in the provided context
+    LLAMA_API const char * llama_token_to_str(struct llama_context * ctx, llama_token token);
+
+    // Special tokens
+    LLAMA_API llama_token llama_token_bos();
+    LLAMA_API llama_token llama_token_eos();
+
+    // TODO: improve the last_n_tokens interface ?
+    LLAMA_API llama_token llama_sample_top_p_top_k(
+       struct llama_context * ctx,
+          const llama_token * last_n_tokens_data,
+                        int   last_n_tokens_size,
+                        int   top_k,
+                      float   top_p,
+                      float   temp,
+                      float   repeat_penalty);
+
+    // Performance information
+    LLAMA_API void llama_print_timings(struct llama_context * ctx);
+    LLAMA_API void llama_reset_timings(struct llama_context * ctx);
+
+    // Print system information
+    LLAMA_API const char * llama_print_system_info(void);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif // LLAMA_H

examples/talk-llama/llama_internal.h

@@ -0,0 +1,12 @@
+// Internal header to be included by llama.cpp and tests/benchmarks only.
+
+#ifndef LLAMA_INTERNAL_H
+#define LLAMA_INTERNAL_H
+
+#include <vector>
+#include <string>
+struct ggml_tensor;
+
+std::vector<std::pair<std::string, struct ggml_tensor *>>& llama_internal_get_tensor_map(struct llama_context * ctx);
+
+#endif // LLAMA_INTERNAL_H

examples/talk-llama/llama_util.h

@@ -0,0 +1,383 @@
+// Internal header to be included only by llama.cpp.
+// Contains wrappers around OS interfaces.
+
+#ifndef LLAMA_UTIL_H
+#define LLAMA_UTIL_H
+
+#include <cstdio>
+#include <cstdint>
+#include <cerrno>
+#include <cstring>
+#include <cstdarg>
+#include <cstdlib>
+#include <climits>
+
+#include <string>
+#include <vector>
+
+#ifdef __has_include
+    #if __has_include(<unistd.h>)
+        #include <unistd.h>
+        #if defined(_POSIX_MAPPED_FILES)
+            #include <sys/mman.h>
+        #endif
+    #endif
+#endif
+
+#if defined(_WIN32)
+    #define WIN32_LEAN_AND_MEAN
+    #define NOMINMAX
+    #include <windows.h>
+    #include <io.h>
+    #include <stdio.h> // for _fseeki64
+#endif
+
+#define LLAMA_ASSERT(x) \
+    do { \
+        if (!(x)) { \
+            fprintf(stderr, "LLAMA_ASSERT: %s:%d: %s\n", __FILE__, __LINE__, #x); \
+            abort(); \
+        } \
+    } while (0)
+
+#ifdef __GNUC__
+__attribute__((format(printf, 1, 2)))
+#endif
+static std::string format(const char * fmt, ...) {
+    va_list ap, ap2;
+    va_start(ap, fmt);
+    va_copy(ap2, ap);
+    int size = vsnprintf(NULL, 0, fmt, ap);
+    LLAMA_ASSERT(size >= 0 && size < INT_MAX);
+    std::vector<char> buf(size + 1);
+    int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
+    LLAMA_ASSERT(size2 == size);
+    va_end(ap2);
+    va_end(ap);
+    return std::string(buf.data(), size);
+};
+
+struct llama_file {
+    // use FILE * so we don't have to re-open the file to mmap
+    FILE * fp;
+    size_t size;
+
+    llama_file(const char * fname, const char * mode) {
+        fp = std::fopen(fname, mode);
+        if (fp == NULL) {
+            throw format("failed to open %s: %s", fname, std::strerror(errno));
+        }
+        seek(0, SEEK_END);
+        size = tell();
+        seek(0, SEEK_SET);
+    }
+
+    size_t tell() const {
+#ifdef _WIN32
+        __int64 ret = _ftelli64(fp);
+#else
+        long ret = std::ftell(fp);
+#endif
+        LLAMA_ASSERT(ret != -1); // this really shouldn't fail
+        return (size_t) ret;
+    }
+
+    void seek(size_t offset, int whence) {
+#ifdef _WIN32
+        int ret = _fseeki64(fp, (__int64) offset, whence);
+#else
+        int ret = std::fseek(fp, (long) offset, whence);
+#endif
+        LLAMA_ASSERT(ret == 0); // same
+    }
+
+    void read_raw(void * ptr, size_t size) {
+        if (size == 0) {
+            return;
+        }
+        errno = 0;
+        std::size_t ret = std::fread(ptr, size, 1, fp);
+        if (ferror(fp)) {
+            throw format("read error: %s", strerror(errno));
+        }
+        if (ret != 1) {
+            throw std::string("unexpectedly reached end of file");
+        }
+    }
+
+    std::uint32_t read_u32() {
+        std::uint32_t ret;
+        read_raw(&ret, sizeof(ret));
+        return ret;
+    }
+
+    std::string read_string(std::uint32_t len) {
+        std::vector<char> chars(len);
+        read_raw(chars.data(), len);
+        return std::string(chars.data(), len);
+    }
+
+    void write_raw(const void * ptr, size_t size) {
+        if (size == 0) {
+            return;
+        }
+        errno = 0;
+        size_t ret = std::fwrite(ptr, size, 1, fp);
+        if (ret != 1) {
+            throw format("write error: %s", strerror(errno));
+        }
+    }
+
+    void write_u32(std::uint32_t val) {
+        write_raw(&val, sizeof(val));
+    }
+
+    ~llama_file() {
+        if (fp) {
+            std::fclose(fp);
+        }
+    }
+};
+
+#if defined(_WIN32)
+static std::string llama_format_win_err(DWORD err) {
+    LPSTR buf;
+    size_t size = FormatMessageA(FORMAT_MESSAGE_ALLOCATE_BUFFER | FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS,
+                                 NULL, err, MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), (LPSTR)&buf, 0, NULL);
+    if (!size) {
+        return "FormatMessageA failed";
+    }
+    std::string ret(buf, size);
+    LocalFree(buf);
+    return ret;
+}
+#endif
+
+struct llama_mmap {
+    void * addr;
+    size_t size;
+
+    llama_mmap(const llama_mmap &) = delete;
+
+#ifdef _POSIX_MAPPED_FILES
+    static constexpr bool SUPPORTED = true;
+
+    llama_mmap(struct llama_file * file) {
+        size = file->size;
+        int fd = fileno(file->fp);
+        int flags = MAP_SHARED;
+#ifdef __linux__
+        flags |= MAP_POPULATE;
+#endif
+        addr = mmap(NULL, file->size, PROT_READ, flags, fd, 0);
+        close(fd);
+        if (addr == MAP_FAILED) {
+            throw format("mmap failed: %s", strerror(errno));
+        }
+
+        // Advise the kernel to preload the mapped memory
+        if (madvise(addr, file->size, MADV_WILLNEED)) {
+            fprintf(stderr, "warning: madvise(.., MADV_WILLNEED) failed: %s\n",
+                    strerror(errno));
+        }
+    }
+
+    ~llama_mmap() {
+        munmap(addr, size);
+    }
+#elif defined(_WIN32)
+    static constexpr bool SUPPORTED = true;
+
+    llama_mmap(struct llama_file * file) {
+        size = file->size;
+
+        HANDLE hFile = (HANDLE) _get_osfhandle(_fileno(file->fp));
+
+        HANDLE hMapping = CreateFileMappingA(hFile, NULL, PAGE_READONLY, 0, 0, NULL);
+        DWORD error = GetLastError();
+        CloseHandle(hFile);
+
+        if (hMapping == NULL) {
+            throw format("CreateFileMappingA failed: %s", llama_format_win_err(error).c_str());
+        }
+
+        addr = MapViewOfFile(hMapping, FILE_MAP_READ, 0, 0, 0);
+        error = GetLastError();
+        CloseHandle(hMapping);
+
+        if (addr == NULL) {
+            throw format("MapViewOfFile failed: %s", llama_format_win_err(error).c_str());
+        }
+
+        // Advise the kernel to preload the mapped memory
+        WIN32_MEMORY_RANGE_ENTRY range;
+        range.VirtualAddress = addr;
+        range.NumberOfBytes = (SIZE_T)size;
+        if (!PrefetchVirtualMemory(GetCurrentProcess(), 1, &range, 0)) {
+            fprintf(stderr, "warning: PrefetchVirtualMemory failed: %s\n",
+                    llama_format_win_err(GetLastError()).c_str());
+        }
+    }
+
+    ~llama_mmap() {
+        if (!UnmapViewOfFile(addr)) {
+            fprintf(stderr, "warning: UnmapViewOfFile failed: %s\n",
+                    llama_format_win_err(GetLastError()).c_str());
+        }
+    }
+#else
+    static constexpr bool SUPPORTED = false;
+
+    llama_mmap(struct llama_file *) {
+        throw std::string("mmap not supported");
+    }
+#endif
+};
+
+// Represents some region of memory being locked using mlock or VirtualLock;
+// will automatically unlock on destruction.
+struct llama_mlock {
+    void * addr = NULL;
+    size_t size = 0;
+    bool failed_already = false;
+
+    llama_mlock() {}
+    llama_mlock(const llama_mlock &) = delete;
+
+    ~llama_mlock() {
+        if (size) {
+            raw_unlock(addr, size);
+        }
+    }
+
+    void init(void * addr) {
+        LLAMA_ASSERT(this->addr == NULL && this->size == 0);
+        this->addr = addr;
+    }
+
+    void grow_to(size_t target_size) {
+        LLAMA_ASSERT(addr);
+        if (failed_already) {
+            return;
+        }
+        size_t granularity = lock_granularity();
+        target_size = (target_size + granularity - 1) & ~(granularity - 1);
+        if (target_size > size) {
+            if (raw_lock((uint8_t *) addr + size, target_size - size)) {
+                size = target_size;
+            } else {
+                failed_already = true;
+            }
+        }
+    }
+
+#ifdef _POSIX_MEMLOCK_RANGE
+    static constexpr bool SUPPORTED = true;
+
+    size_t lock_granularity() {
+        return (size_t) sysconf(_SC_PAGESIZE);
+    }
+
+    #ifdef __APPLE__
+        #define MLOCK_SUGGESTION \
+            "Try increasing the sysctl values 'vm.user_wire_limit' and 'vm.global_user_wire_limit' and/or " \
+            "decreasing 'vm.global_no_user_wire_amount'.  Also try increasing RLIMIT_MLOCK (ulimit -l).\n"
+    #else
+        #define MLOCK_SUGGESTION \
+            "Try increasing RLIMIT_MLOCK ('ulimit -l' as root).\n"
+    #endif
+
+    bool raw_lock(const void * addr, size_t size) {
+        if (!mlock(addr, size)) {
+            return true;
+        } else {
+            fprintf(stderr, "warning: failed to mlock %zu-byte buffer (after previously locking %zu bytes): %s\n" MLOCK_SUGGESTION,
+                    size, this->size, std::strerror(errno));
+            return false;
+        }
+    }
+
+    #undef MLOCK_SUGGESTION
+
+    void raw_unlock(void * addr, size_t size) {
+        if (munlock(addr, size)) {
+            fprintf(stderr, "warning: failed to munlock buffer: %s\n", std::strerror(errno));
+        }
+    }
+#elif defined(_WIN32)
+    static constexpr bool SUPPORTED = true;
+
+    size_t lock_granularity() {
+        SYSTEM_INFO si;
+        GetSystemInfo(&si);
+        return (size_t) si.dwPageSize;
+    }
+
+    bool raw_lock(void * addr, size_t size) {
+        for (int tries = 1; ; tries++) {
+            if (VirtualLock(addr, size)) {
+                return true;
+            }
+            if (tries == 2) {
+                fprintf(stderr, "warning: failed to VirtualLock %zu-byte buffer (after previously locking %zu bytes): %s\n",
+                        size, this->size, llama_format_win_err(GetLastError()).c_str());
+                return false;
+            }
+
+            // It failed but this was only the first try; increase the working
+            // set size and try again.
+            SIZE_T min_ws_size, max_ws_size;
+            if (!GetProcessWorkingSetSize(GetCurrentProcess(), &min_ws_size, &max_ws_size)) {
+                fprintf(stderr, "warning: GetProcessWorkingSetSize failed: %s\n",
+                        llama_format_win_err(GetLastError()).c_str());
+                return false;
+            }
+            // Per MSDN: "The maximum number of pages that a process can lock
+            // is equal to the number of pages in its minimum working set minus
+            // a small overhead."
+            // Hopefully a megabyte is enough overhead:
+            size_t increment = size + 1048576;
+            // The minimum must be <= the maximum, so we need to increase both:
+            min_ws_size += size;
+            max_ws_size += size;
+            if (!SetProcessWorkingSetSize(GetCurrentProcess(), min_ws_size, max_ws_size)) {
+                fprintf(stderr, "warning: SetProcessWorkingSetSize failed: %s\n",
+                        llama_format_win_err(GetLastError()).c_str());
+                return false;
+            }
+        }
+    }
+
+    void raw_unlock(void * addr, size_t size) {
+        if (!VirtualUnlock(addr, size)) {
+            fprintf(stderr, "warning: failed to VirtualUnlock buffer: %s\n",
+                    llama_format_win_err(GetLastError()).c_str());
+        }
+    }
+#else
+    static constexpr bool SUPPORTED = false;
+
+    void raw_lock(const void * addr, size_t size) {
+        fprintf(stderr, "warning: mlock not supported on this system\n");
+    }
+
+    void raw_unlock(const void * addr, size_t size) {}
+#endif
+};
+
+// Replacement for std::vector<uint8_t> that doesn't require zero-initialization.
+struct llama_buffer {
+    uint8_t * addr = NULL;
+    size_t size = 0;
+
+    void resize(size_t size) {
+        delete[] addr;
+        addr = new uint8_t[size];
+        this->size = size;
+    }
+
+    ~llama_buffer() {
+        delete[] addr;
+    }
+};
+#endif

examples/talk-llama/prompts/talk-alpaca.txt

@@ -0,0 +1,23 @@
+Below is an instruction that describes a task. Write a response that appropriately completes the request.
+
+### Instruction:
+
+Write a text transcript of a never ending dialog, where {0} interacts with an AI assistant named {1}.
+{1} is helpful, kind, honest, friendly, good at writing and never fails to answer {0}’s requests immediately and with details and precision.
+There are no annotations like (30 seconds passed...) or (to himself), just what {0} and {1} say aloud to each other.
+The transcript only includes text, it does not include markup like HTML and Markdown.
+{1} responds with short and concise answers.
+
+### Response:
+
+{0}{4} Hello, {1}!
+{1}{4} Hello {0}! How may I help you today?
+{0}{4} What time is it?
+{1}{4} It is {2} o'clock.
+{0}{4} What year is it?
+{1}{4} We are in {3}.
+{0}{4} What is a cat?
+{1}{4} A cat is a domestic species of small carnivorous mammal. It is the only domesticated species in the family Felidae.
+{0}{4} Name a color.
+{1}{4} Blue
+{0}{4}

examples/talk-llama/speak.sh

@@ -0,0 +1,21 @@
+#!/bin/bash
+
+# Usage:
+#  speak.sh <voice_id> <text-to-speak>
+
+# espeak
+# Mac OS: brew install espeak
+# Linux: apt-get install espeak
+#
+#espeak -v en-us+m$1 -s 225 -p 50 -a 200 -g 5 -k 5 "$2"
+
+# for Mac
+say "$2"
+
+# Eleven Labs
+# To use it, install the elevenlabs module from pip (pip install elevenlabs), register to https://beta.elevenlabs.io to get an api key and paste it in /examples/talk-llama/eleven-labs.py 
+#
+#wd=$(dirname $0)
+#script=$wd/eleven-labs.py
+#python3 $script $1 "$2" >/dev/null 2>&1
+#ffplay -autoexit -nodisp -loglevel quiet -hide_banner -i ./audio.mp3 >/dev/null 2>&1

examples/talk-llama/talk-llama.cpp

@@ -0,0 +1,550 @@
+// Talk with AI
+//
+
+#include "common.h"
+#include "common-sdl.h"
+#include "whisper.h"
+#include "llama.h"
+
+#include <cassert>
+#include <cstdio>
+#include <fstream>
+#include <regex>
+#include <string>
+#include <thread>
+#include <vector>
+#include <regex>
+
+std::vector<llama_token> llama_tokenize(struct llama_context * ctx, const std::string & text, bool add_bos) {
+    // initialize to prompt numer of chars, since n_tokens <= n_prompt_chars
+    std::vector<llama_token> res(text.size() + (int)add_bos);
+    int n = llama_tokenize(ctx, text.c_str(), res.data(), res.size(), add_bos);
+    assert(n >= 0);
+    res.resize(n);
+
+    return res;
+}
+
+// command-line parameters
+struct whisper_params {
+    int32_t n_threads  = std::min(4, (int32_t) std::thread::hardware_concurrency());
+    int32_t voice_ms   = 10000;
+    int32_t capture_id = -1;
+    int32_t max_tokens = 32;
+    int32_t audio_ctx  = 0;
+
+    int32_t n_parts_llama = -1;
+
+    float vad_thold    = 0.6f;
+    float freq_thold   = 100.0f;
+
+    bool speed_up      = false;
+    bool translate     = false;
+    bool print_special = false;
+    bool print_energy  = false;
+    bool no_timestamps = true;
+    bool verbose_prompt = false;
+
+    std::string person      = "Georgi";
+    std::string language    = "en";
+    std::string model_wsp   = "models/ggml-base.en.bin";
+    std::string model_llama = "models/ggml-llama-7B.bin";
+    std::string speak       = "./examples/talk-llama/speak.sh";
+    std::string prompt      = "";
+    std::string fname_out;
+};
+
+void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
+
+bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
+    for (int i = 1; i < argc; i++) {
+        std::string arg = argv[i];
+
+        if (arg == "-h" || arg == "--help") {
+            whisper_print_usage(argc, argv, params);
+            exit(0);
+        }
+        else if (arg == "-t"   || arg == "--threads")       { params.n_threads     = std::stoi(argv[++i]); }
+        else if (arg == "-vms" || arg == "--voice-ms")      { params.voice_ms      = std::stoi(argv[++i]); }
+        else if (arg == "-c"   || arg == "--capture")       { params.capture_id    = std::stoi(argv[++i]); }
+        else if (arg == "-mt"  || arg == "--max-tokens")    { params.max_tokens    = std::stoi(argv[++i]); }
+        else if (arg == "-ac"  || arg == "--audio-ctx")     { params.audio_ctx     = std::stoi(argv[++i]); }
+        else if (arg == "-vth" || arg == "--vad-thold")     { params.vad_thold     = std::stof(argv[++i]); }
+        else if (arg == "-fth" || arg == "--freq-thold")    { params.freq_thold    = std::stof(argv[++i]); }
+        else if (arg == "--n-parts-llama")                  { params.n_parts_llama = std::stoi(argv[++i]); }
+        else if (arg == "-su"  || arg == "--speed-up")      { params.speed_up      = true; }
+        else if (arg == "-tr"  || arg == "--translate")     { params.translate     = true; }
+        else if (arg == "-ps"  || arg == "--print-special") { params.print_special = true; }
+        else if (arg == "-pe"  || arg == "--print-energy")  { params.print_energy  = true; }
+        else if (arg == "--verbose-prompt")                 { params.verbose_prompt = true; }
+        else if (arg == "-p"   || arg == "--person")        { params.person        = argv[++i]; }
+        else if (arg == "-l"   || arg == "--language")      { params.language      = argv[++i]; }
+        else if (arg == "-mw"  || arg == "--model-whisper") { params.model_wsp     = argv[++i]; }
+        else if (arg == "-ml"  || arg == "--model-llama")   { params.model_llama   = argv[++i]; }
+        else if (arg == "-s"   || arg == "--speak")         { params.speak         = argv[++i]; }
+        else if (arg == "--prompt-file")                    {
+            std::ifstream file(argv[++i]);
+            std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(params.prompt));
+            if (params.prompt.back() == '\n') {
+                params.prompt.pop_back();
+            }
+        }
+        else if (arg == "-f"   || arg == "--file")          { params.fname_out     = argv[++i]; }
+        else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            whisper_print_usage(argc, argv, params);
+            exit(0);
+        }
+    }
+
+    return true;
+}
+
+void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & params) {
+    fprintf(stderr, "\n");
+    fprintf(stderr, "usage: %s [options]\n", argv[0]);
+    fprintf(stderr, "\n");
+    fprintf(stderr, "options:\n");
+    fprintf(stderr, "  -h,       --help          [default] show this help message and exit\n");
+    fprintf(stderr, "  -t N,     --threads N     [%-7d] number of threads to use during computation\n", params.n_threads);
+    fprintf(stderr, "  -vms N,   --voice-ms N    [%-7d] voice duration in milliseconds\n",              params.voice_ms);
+    fprintf(stderr, "  -c ID,    --capture ID    [%-7d] capture device ID\n",                           params.capture_id);
+    fprintf(stderr, "  -mt N,    --max-tokens N  [%-7d] maximum number of tokens per audio chunk\n",    params.max_tokens);
+    fprintf(stderr, "  -ac N,    --audio-ctx N   [%-7d] audio context size (0 - all)\n",                params.audio_ctx);
+    fprintf(stderr, "  -vth N,   --vad-thold N   [%-7.2f] voice activity detection threshold\n",        params.vad_thold);
+    fprintf(stderr, "  -fth N,   --freq-thold N  [%-7.2f] high-pass frequency cutoff\n",                params.freq_thold);
+    fprintf(stderr, "  -su,      --speed-up      [%-7s] speed up audio by x2 (reduced accuracy)\n",     params.speed_up ? "true" : "false");
+    fprintf(stderr, "  -tr,      --translate     [%-7s] translate from source language to english\n",   params.translate ? "true" : "false");
+    fprintf(stderr, "  -ps,      --print-special [%-7s] print special tokens\n",                        params.print_special ? "true" : "false");
+    fprintf(stderr, "  -pe,      --print-energy  [%-7s] print sound energy (for debugging)\n",          params.print_energy ? "true" : "false");
+    fprintf(stderr, "  -p NAME,  --person NAME   [%-7s] person name (for prompt selection)\n",          params.person.c_str());
+    fprintf(stderr, "  -l LANG,  --language LANG [%-7s] spoken language\n",                             params.language.c_str());
+    fprintf(stderr, "  -mw FILE, --model-whisper [%-7s] whisper model file\n",                          params.model_wsp.c_str());
+    fprintf(stderr, "  -ml FILE, --model-llama   [%-7s] llama model file\n",                            params.model_llama.c_str());
+    fprintf(stderr, "  --n-parts-llama N         [%-7d] num parts in llama model file\n",               params.n_parts_llama);
+    fprintf(stderr, "  -s FILE,  --speak TEXT    [%-7s] command for TTS\n",                             params.speak.c_str());
+    fprintf(stderr, "  --prompt-file FNAME       [%-7s] file with custom prompt to start dialog\n",     "");
+    fprintf(stderr, "  --verbose-prompt          [%-7s] print prompt at start\n",                       params.verbose_prompt ? "true" : "false");
+    fprintf(stderr, "  -f FNAME, --file FNAME    [%-7s] text output file name\n",                       params.fname_out.c_str());
+    fprintf(stderr, "\n");
+}
+
+std::string transcribe(
+        whisper_context * ctx,
+        const whisper_params & params,
+        const std::vector<float> & pcmf32,
+        const std::string prompt_text,
+        float & prob,
+        int64_t & t_ms) {
+    const auto t_start = std::chrono::high_resolution_clock::now();
+
+    prob = 0.0f;
+    t_ms = 0;
+
+    std::vector<whisper_token> prompt_tokens;
+
+    whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
+
+    prompt_tokens.resize(1024);
+    prompt_tokens.resize(whisper_tokenize(ctx, prompt_text.c_str(), prompt_tokens.data(), prompt_tokens.size()));
+
+    wparams.print_progress   = false;
+    wparams.print_special    = params.print_special;
+    wparams.print_realtime   = false;
+    wparams.print_timestamps = !params.no_timestamps;
+    wparams.translate        = params.translate;
+    wparams.no_context       = true;
+    wparams.single_segment   = true;
+    wparams.max_tokens       = params.max_tokens;
+    wparams.language         = params.language.c_str();
+    wparams.n_threads        = params.n_threads;
+
+    wparams.prompt_tokens    = prompt_tokens.empty() ? nullptr : prompt_tokens.data();
+    wparams.prompt_n_tokens  = prompt_tokens.empty() ? 0       : prompt_tokens.size();
+
+    wparams.audio_ctx        = params.audio_ctx;
+    wparams.speed_up         = params.speed_up;
+
+    if (whisper_full(ctx, wparams, pcmf32.data(), pcmf32.size()) != 0) {
+        return "";
+    }
+
+    int prob_n = 0;
+    std::string result;
+
+    const int n_segments = whisper_full_n_segments(ctx);
+    for (int i = 0; i < n_segments; ++i) {
+        const char * text = whisper_full_get_segment_text(ctx, i);
+
+        result += text;
+
+        const int n_tokens = whisper_full_n_tokens(ctx, i);
+        for (int j = 0; j < n_tokens; ++j) {
+            const auto token = whisper_full_get_token_data(ctx, i, j);
+
+            prob += token.p;
+            ++prob_n;
+        }
+    }
+
+    if (prob_n > 0) {
+        prob /= prob_n;
+    }
+
+    const auto t_end = std::chrono::high_resolution_clock::now();
+    t_ms = std::chrono::duration_cast<std::chrono::milliseconds>(t_end - t_start).count();
+
+    return result;
+}
+
+const std::string k_prompt_whisper = R"(A conversation with a person called {1}.)";
+
+const std::string k_prompt_llama = R"(Text transcript of a never ending dialog, where {0} interacts with an AI assistant named {1}.
+{1} is helpful, kind, honest, friendly, good at writing and never fails to answer {0}’s requests immediately and with details and precision.
+There are no annotations like (30 seconds passed...) or (to himself), just what {0} and {1} say aloud to each other.
+The transcript only includes text, it does not include markup like HTML and Markdown.
+{1} responds with short and concise answers.
+
+{0}{4} Hello, {1}!
+{1}{4} Hello {0}! How may I help you today?
+{0}{4} What time is it?
+{1}{4} It is {2} o'clock.
+{0}{4} What year is it?
+{1}{4} We are in {3}.
+{0}{4} What is a cat?
+{1}{4} A cat is a domestic species of small carnivorous mammal. It is the only domesticated species in the family Felidae.
+{0}{4} Name a color.
+{1}{4} Blue
+{0}{4})";
+
+int main(int argc, char ** argv) {
+    whisper_params params;
+
+    if (whisper_params_parse(argc, argv, params) == false) {
+        return 1;
+    }
+
+    if (whisper_lang_id(params.language.c_str()) == -1) {
+        fprintf(stderr, "error: unknown language '%s'\n", params.language.c_str());
+        whisper_print_usage(argc, argv, params);
+        exit(0);
+    }
+
+    // whisper init
+
+    struct whisper_context * ctx_wsp = whisper_init_from_file(params.model_wsp.c_str());
+
+    // llama init
+
+    auto lparams = llama_context_default_params();
+
+    // tune these to your liking
+    lparams.n_ctx      = 2048;
+    lparams.seed       = 1;
+    lparams.f16_kv     = true;
+    lparams.n_parts    = params.n_parts_llama;
+
+    struct llama_context * ctx_llama = llama_init_from_file(params.model_llama.c_str(), lparams);
+
+    // print some info about the processing
+    {
+        fprintf(stderr, "\n");
+
+        if (!whisper_is_multilingual(ctx_wsp)) {
+            if (params.language != "en" || params.translate) {
+                params.language = "en";
+                params.translate = false;
+                fprintf(stderr, "%s: WARNING: model is not multilingual, ignoring language and translation options\n", __func__);
+            }
+        }
+        fprintf(stderr, "%s: processing, %d threads, lang = %s, task = %s, timestamps = %d ...\n",
+                __func__,
+                params.n_threads,
+                params.language.c_str(),
+                params.translate ? "translate" : "transcribe",
+                params.no_timestamps ? 0 : 1);
+
+        fprintf(stderr, "\n");
+    }
+
+
+    // init audio
+
+    audio_async audio(30*1000);
+    if (!audio.init(params.capture_id, WHISPER_SAMPLE_RATE)) {
+        fprintf(stderr, "%s: audio.init() failed!\n", __func__);
+        return 1;
+    }
+
+    audio.resume();
+
+    int n_iter = 0;
+
+    bool is_running  = true;
+    bool force_speak = false;
+
+    float prob0 = 0.0f;
+
+    const std::string chat_symb = ":";
+    const std::string bot_name  = "LLaMA";
+
+    std::vector<float> pcmf32_cur;
+    std::vector<float> pcmf32_prompt;
+
+    const std::string prompt_whisper = ::replace(k_prompt_whisper, "{1}", bot_name);
+
+    // construct the initial prompt for LLaMA inference
+    std::string prompt_llama = params.prompt.empty() ? k_prompt_llama : params.prompt;
+
+    // need to have leading ' '
+    prompt_llama.insert(0, 1, ' ');
+
+    prompt_llama = ::replace(prompt_llama, "{0}", params.person);
+    prompt_llama = ::replace(prompt_llama, "{1}", bot_name);
+
+    {
+        // get time string
+        std::string time_str;
+        {
+            time_t t = time(0);
+            struct tm * now = localtime(&t);
+            char buf[128];
+            strftime(buf, sizeof(buf), "%H:%M", now);
+            time_str = buf;
+        }
+        prompt_llama = ::replace(prompt_llama, "{2}", time_str);
+    }
+
+    {
+        // get year string
+        std::string year_str;
+        {
+            time_t t = time(0);
+            struct tm * now = localtime(&t);
+            char buf[128];
+            strftime(buf, sizeof(buf), "%Y", now);
+            year_str = buf;
+        }
+        prompt_llama = ::replace(prompt_llama, "{3}", year_str);
+    }
+
+    prompt_llama = ::replace(prompt_llama, "{4}", chat_symb);
+
+    // evaluate the initial prompt
+
+    auto embd_inp = ::llama_tokenize(ctx_llama, prompt_llama, true);
+
+    printf("\n");
+    printf("%s : initializing - please wait ...\n", __func__);
+
+    if (llama_eval(ctx_llama, embd_inp.data(), embd_inp.size(), 0, params.n_threads)) {
+        fprintf(stderr, "%s : failed to eval\n", __func__);
+        return 1;
+    }
+
+    if (params.verbose_prompt) {
+        fprintf(stdout, "\n");
+        fprintf(stdout, "%s", prompt_llama.c_str());
+        fflush(stdout);
+    }
+
+    printf("%s : done! start speaking in the microphone\n", __func__);
+    printf("\n");
+    printf("%s%s", params.person.c_str(), chat_symb.c_str());
+    fflush(stdout);
+
+    // clear audio buffer
+    audio.clear();
+
+    // text inference variables
+    const int voice_id = 2;
+    const int n_keep   = embd_inp.size();
+    const int n_ctx    = llama_n_ctx(ctx_llama);
+
+    int n_past = n_keep;
+    int n_prev = 64; // TODO arg
+
+    std::vector<llama_token> embd;
+
+    // reverse prompts for detecting when it's time to stop speaking
+    std::vector<std::string> antiprompts = {
+        params.person + chat_symb,
+    };
+
+    // main loop
+    while (is_running) {
+        // handle Ctrl + C
+        is_running = sdl_poll_events();
+
+        if (!is_running) {
+            break;
+        }
+
+        // delay
+        std::this_thread::sleep_for(std::chrono::milliseconds(100));
+
+        int64_t t_ms = 0;
+
+        {
+            audio.get(2000, pcmf32_cur);
+
+            if (::vad_simple(pcmf32_cur, WHISPER_SAMPLE_RATE, 1250, params.vad_thold, params.freq_thold, params.print_energy) || force_speak) {
+                //fprintf(stdout, "%s: Speech detected! Processing ...\n", __func__);
+
+                audio.get(params.voice_ms, pcmf32_cur);
+
+                std::string text_heard;
+
+                if (!force_speak) {
+                    text_heard = ::trim(::transcribe(ctx_wsp, params, pcmf32_cur, prompt_whisper, prob0, t_ms));
+                }
+
+                // remove text between brackets using regex
+                {
+                    std::regex re("\\[.*?\\]");
+                    text_heard = std::regex_replace(text_heard, re, "");
+                }
+
+                // remove text between brackets using regex
+                {
+                    std::regex re("\\(.*?\\)");
+                    text_heard = std::regex_replace(text_heard, re, "");
+                }
+
+                // remove all characters, except for letters, numbers, punctuation and ':', '\'', '-', ' '
+                text_heard = std::regex_replace(text_heard, std::regex("[^a-zA-Z0-9\\.,\\?!\\s\\:\\'\\-]"), "");
+
+                // take first line
+                text_heard = text_heard.substr(0, text_heard.find_first_of('\n'));
+
+                // remove leading and trailing whitespace
+                text_heard = std::regex_replace(text_heard, std::regex("^\\s+"), "");
+                text_heard = std::regex_replace(text_heard, std::regex("\\s+$"), "");
+
+                const std::vector<llama_token> tokens = llama_tokenize(ctx_llama, text_heard.c_str(), false);
+
+                if (text_heard.empty() || tokens.empty() || force_speak) {
+                    //fprintf(stdout, "%s: Heard nothing, skipping ...\n", __func__);
+                    audio.clear();
+
+                    continue;
+                }
+
+                force_speak = false;
+
+                text_heard.insert(0, 1, ' ');
+                text_heard += "\n" + bot_name + chat_symb;
+                fprintf(stdout, "%s%s%s", "\033[1m", text_heard.c_str(), "\033[0m");
+                fflush(stdout);
+
+                embd = ::llama_tokenize(ctx_llama, text_heard, false);
+
+                // text inference
+                bool done = false;
+                std::string text_to_speak;
+                while (true) {
+                    // predict
+                    if (embd.size() > 0) {
+                        if (n_past + (int) embd.size() > n_ctx) {
+                            n_past = n_keep;
+
+                            // insert n_left/2 tokens at the start of embd from last_n_tokens
+                            embd.insert(embd.begin(), embd_inp.begin() + embd_inp.size() - n_prev, embd_inp.end());
+
+                            //printf("\n---\n");
+                            //printf("resetting: '");
+                            //for (int i = 0; i < (int) embd.size(); i++) {
+                            //    printf("%s", llama_token_to_str(ctx_llama, embd[i]));
+                            //}
+                            //printf("'\n");
+                            //printf("\n---\n");
+                        }
+
+                        if (llama_eval(ctx_llama, embd.data(), embd.size(), n_past, params.n_threads)) {
+                            fprintf(stderr, "%s : failed to eval\n", __func__);
+                            return 1;
+                        }
+                    }
+
+                    //printf("n_iter = %d, n_past = %d, n_ctx = %d, n_keep = %d, n_prev = %d, embd.size() = %d\n", n_iter, n_past, n_ctx, n_keep, n_prev, (int) embd.size());
+
+                    embd_inp.insert(embd_inp.end(), embd.begin(), embd.end());
+                    n_past += embd.size();
+                    embd.clear();
+
+                    if (done) break;
+
+                    {
+                        // out of user input, sample next token
+                        const float top_k          = 5;
+                        const float top_p          = 0.80f;
+                        const float temp           = 0.30f;
+                        const float repeat_penalty = 1.1764f;
+
+                        const int repeat_last_n    = 256;
+
+                        llama_token id = 0;
+
+                        {
+                            auto logits = llama_get_logits(ctx_llama);
+                            logits[llama_token_eos()] = 0;
+
+                            id = llama_sample_top_p_top_k(ctx_llama,
+                                    embd_inp.data() + std::max(0, n_past - repeat_last_n),
+                                    repeat_last_n, top_k, top_p, temp, repeat_penalty);
+                        }
+
+                        if (id != llama_token_eos()) {
+                            // add it to the context
+                            embd.push_back(id);
+
+                            text_to_speak += llama_token_to_str(ctx_llama, id);
+
+                            printf("%s", llama_token_to_str(ctx_llama, id));
+                        }
+                    }
+
+                    {
+                        std::string last_output;
+                        for (int i = embd_inp.size() - 16; i < (int) embd_inp.size(); i++) {
+                            last_output += llama_token_to_str(ctx_llama, embd_inp[i]);
+                        }
+                        last_output += llama_token_to_str(ctx_llama, embd[0]);
+
+                        for (std::string & antiprompt : antiprompts) {
+                            if (last_output.find(antiprompt.c_str(), last_output.length() - antiprompt.length(), antiprompt.length()) != std::string::npos) {
+                                done = true;
+                                text_to_speak = ::replace(text_to_speak, antiprompt, "");
+                                fflush(stdout);
+                                break;
+                            }
+                        }
+                    }
+
+                    is_running = sdl_poll_events();
+
+                    if (!is_running) {
+                        break;
+                    }
+                }
+
+                text_to_speak = ::replace(text_to_speak, "\"", "");
+                system((params.speak + " " + std::to_string(voice_id) + " \"" + text_to_speak + "\"").c_str());
+
+                audio.clear();
+
+                ++n_iter;
+            }
+        }
+    }
+
+    audio.pause();
+
+    whisper_print_timings(ctx_wsp);
+    whisper_free(ctx_wsp);
+
+    llama_print_timings(ctx_llama);
+    llama_free(ctx_llama);
+
+    return 0;
+}

examples/talk.wasm/gpt-2.cpp

@@ -841,6 +841,7 @@ struct gpt2_context * gpt2_init(const char * path_model) {
 
         if (!gpt2_model_load(path_model, ctx->model, ctx->vocab)) {
             fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, "gpt-2.bin");
+            delete ctx;
             return nullptr;
         }
 

examples/talk/.gitignore

@@ -1 +1 @@
-eleven-labs.py
+audio.mp3

examples/talk/CMakeLists.txt

@@ -7,7 +7,7 @@ if (WHISPER_SUPPORT_SDL2)
 
     # TODO: this is temporary
     #       need to export ggml symbols for MSVC, but too lazy ..
-    add_executable(${TARGET} talk.cpp gpt-2.cpp ../../ggml.c ../../whisper.cpp)
+    add_executable(${TARGET} talk.cpp gpt-2.cpp ../common.cpp ../common-sdl.cpp ../../ggml.c ../../whisper.cpp)
 
     include(DefaultTargetOptions)
 

examples/talk/README.md

@@ -31,7 +31,7 @@ To run this, you will need a ggml GPT-2 model: [instructions](https://github.com
 Alternatively, you can simply download the smallest ggml GPT-2 117M model (240 MB) like this:
 
 ```
-wget --quiet --show-progress -O models/ggml-gpt-2-117M.bin https://huggingface.co/datasets/ggerganov/ggml/raw/main/ggml-model-gpt-2-117M.bin
+wget --quiet --show-progress -O models/ggml-gpt-2-117M.bin https://huggingface.co/ggerganov/ggml/raw/main/ggml-model-gpt-2-117M.bin
 ```
 
 ## TTS

examples/talk/eleven-labs.py

@@ -0,0 +1,23 @@
+import sys
+import importlib.util
+
+api_key = "" #Write your https://beta.elevenlabs.io api key here
+if not api_key:
+    print("To use elevenlabs you have to register to https://beta.elevenlabs.io and add your elevenlabs api key to examples/talk/eleven-labs.py")
+    sys.exit()
+
+if importlib.util.find_spec("elevenlabs") is None:
+    print("elevenlabs library is not installed, you can install it to your enviroment using 'pip install elevenlabs'")
+    sys.exit()
+
+from elevenlabs import ElevenLabs
+eleven = ElevenLabs(api_key)
+
+# Get a Voice object, by name or UUID
+voice = eleven.voices["Arnold"] #Possible Voices: Adam Antoni Arnold Bella Domi Elli Josh
+
+# Generate the TTS
+audio = voice.generate(str(sys.argv[2:]))
+
+# Save the TTS to a file
+audio.save("audio") 

examples/talk/speak.sh

@@ -7,9 +7,13 @@
 # Mac OS: brew install espeak
 # Linux: apt-get install espeak
 #
-espeak -v en-us+m$1 -s 175 -p 50 -a 200 -g 5 -k 5 "$2"
+#espeak -v en-us+m$1 -s 175 -p 50 -a 200 -g 5 -k 5 "$2"
+
+# Mac OS "say" command
+say "$2"
 
 # Eleven Labs
+# To use it, install the elevenlabs module from pip (pip install elevenlabs), register to https://beta.elevenlabs.io to get an api key and paste it in /examples/talk/eleven-labs.py 
 #
 #wd=$(dirname $0)
 #script=$wd/eleven-labs.py

examples/talk/talk.cpp

@@ -1,16 +1,14 @@
 // Talk with AI
 //
 
+#include "common.h"
+#include "common-sdl.h"
 #include "whisper.h"
 #include "gpt-2.h"
 
-#include <SDL.h>
-#include <SDL_audio.h>
-
 #include <cassert>
 #include <cstdio>
 #include <fstream>
-#include <mutex>
 #include <regex>
 #include <string>
 #include <thread>
@@ -105,320 +103,6 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "\n");
 }
 
-//
-// SDL Audio capture
-//
-
-class audio_async {
-public:
-    audio_async(int len_ms);
-    ~audio_async();
-
-    bool init(int capture_id, int sample_rate);
-
-    // start capturing audio via the provided SDL callback
-    // keep last len_ms seconds of audio in a circular buffer
-    bool resume();
-    bool pause();
-    bool clear();
-
-    // callback to be called by SDL
-    void callback(uint8_t * stream, int len);
-
-    // get audio data from the circular buffer
-    void get(int ms, std::vector<float> & audio);
-
-private:
-    SDL_AudioDeviceID m_dev_id_in = 0;
-
-    int m_len_ms = 0;
-    int m_sample_rate = 0;
-
-    bool       m_running = false;
-    std::mutex m_mutex;
-
-    std::vector<float> m_audio;
-    std::vector<float> m_audio_new;
-    size_t             m_audio_pos = 0;
-    size_t             m_audio_len = 0;
-};
-
-audio_async::audio_async(int len_ms) {
-    m_len_ms = len_ms;
-}
-
-audio_async::~audio_async() {
-    if (m_dev_id_in) {
-        SDL_CloseAudioDevice(m_dev_id_in);
-    }
-}
-
-bool audio_async::init(int capture_id, int sample_rate) {
-    SDL_LogSetPriority(SDL_LOG_CATEGORY_APPLICATION, SDL_LOG_PRIORITY_INFO);
-
-    if (SDL_Init(SDL_INIT_AUDIO) < 0) {
-        SDL_LogError(SDL_LOG_CATEGORY_APPLICATION, "Couldn't initialize SDL: %s\n", SDL_GetError());
-        return false;
-    }
-
-    SDL_SetHintWithPriority(SDL_HINT_AUDIO_RESAMPLING_MODE, "medium", SDL_HINT_OVERRIDE);
-
-    {
-        int nDevices = SDL_GetNumAudioDevices(SDL_TRUE);
-        fprintf(stderr, "%s: found %d capture devices:\n", __func__, nDevices);
-        for (int i = 0; i < nDevices; i++) {
-            fprintf(stderr, "%s:    - Capture device #%d: '%s'\n", __func__, i, SDL_GetAudioDeviceName(i, SDL_TRUE));
-        }
-    }
-
-    SDL_AudioSpec capture_spec_requested;
-    SDL_AudioSpec capture_spec_obtained;
-
-    SDL_zero(capture_spec_requested);
-    SDL_zero(capture_spec_obtained);
-
-    capture_spec_requested.freq     = sample_rate;
-    capture_spec_requested.format   = AUDIO_F32;
-    capture_spec_requested.channels = 1;
-    capture_spec_requested.samples  = 1024;
-    capture_spec_requested.callback = [](void * userdata, uint8_t * stream, int len) {
-        audio_async * audio = (audio_async *) userdata;
-        audio->callback(stream, len);
-    };
-    capture_spec_requested.userdata = this;
-
-    if (capture_id >= 0) {
-        fprintf(stderr, "%s: attempt to open capture device %d : '%s' ...\n", __func__, capture_id, SDL_GetAudioDeviceName(capture_id, SDL_TRUE));
-        m_dev_id_in = SDL_OpenAudioDevice(SDL_GetAudioDeviceName(capture_id, SDL_TRUE), SDL_TRUE, &capture_spec_requested, &capture_spec_obtained, 0);
-    } else {
-        fprintf(stderr, "%s: attempt to open default capture device ...\n", __func__);
-        m_dev_id_in = SDL_OpenAudioDevice(nullptr, SDL_TRUE, &capture_spec_requested, &capture_spec_obtained, 0);
-    }
-
-    if (!m_dev_id_in) {
-        fprintf(stderr, "%s: couldn't open an audio device for capture: %s!\n", __func__, SDL_GetError());
-        m_dev_id_in = 0;
-
-        return false;
-    } else {
-        fprintf(stderr, "%s: obtained spec for input device (SDL Id = %d):\n", __func__, m_dev_id_in);
-        fprintf(stderr, "%s:     - sample rate:       %d\n",                   __func__, capture_spec_obtained.freq);
-        fprintf(stderr, "%s:     - format:            %d (required: %d)\n",    __func__, capture_spec_obtained.format,
-                capture_spec_requested.format);
-        fprintf(stderr, "%s:     - channels:          %d (required: %d)\n",    __func__, capture_spec_obtained.channels,
-                capture_spec_requested.channels);
-        fprintf(stderr, "%s:     - samples per frame: %d\n",                   __func__, capture_spec_obtained.samples);
-        fprintf(stderr, "\n");
-    }
-
-    m_sample_rate = capture_spec_obtained.freq;
-
-    m_audio.resize((m_sample_rate*m_len_ms)/1000);
-
-    return true;
-}
-
-bool audio_async::resume() {
-    if (!m_dev_id_in) {
-        fprintf(stderr, "%s: no audio device to resume!\n", __func__);
-        return false;
-    }
-
-    if (m_running) {
-        fprintf(stderr, "%s: already running!\n", __func__);
-        return false;
-    }
-
-    SDL_PauseAudioDevice(m_dev_id_in, 0);
-
-    m_running = true;
-
-    return true;
-}
-
-bool audio_async::pause() {
-    if (!m_dev_id_in) {
-        fprintf(stderr, "%s: no audio device to pause!\n", __func__);
-        return false;
-    }
-
-    if (!m_running) {
-        fprintf(stderr, "%s: already paused!\n", __func__);
-        return false;
-    }
-
-    SDL_PauseAudioDevice(m_dev_id_in, 1);
-
-    m_running = false;
-
-    return true;
-}
-
-bool audio_async::clear() {
-    if (!m_dev_id_in) {
-        fprintf(stderr, "%s: no audio device to clear!\n", __func__);
-        return false;
-    }
-
-    if (!m_running) {
-        fprintf(stderr, "%s: not running!\n", __func__);
-        return false;
-    }
-
-    {
-        std::lock_guard<std::mutex> lock(m_mutex);
-
-        m_audio_pos = 0;
-        m_audio_len = 0;
-    }
-
-    return true;
-}
-
-// callback to be called by SDL
-void audio_async::callback(uint8_t * stream, int len) {
-    if (!m_running) {
-        return;
-    }
-
-    const size_t n_samples = len / sizeof(float);
-
-    m_audio_new.resize(n_samples);
-    memcpy(m_audio_new.data(), stream, n_samples * sizeof(float));
-
-    //fprintf(stderr, "%s: %zu samples, pos %zu, len %zu\n", __func__, n_samples, m_audio_pos, m_audio_len);
-
-    {
-        std::lock_guard<std::mutex> lock(m_mutex);
-
-        if (m_audio_pos + n_samples > m_audio.size()) {
-            const size_t n0 = m_audio.size() - m_audio_pos;
-
-            memcpy(&m_audio[m_audio_pos], stream, n0 * sizeof(float));
-            memcpy(&m_audio[0], &stream[n0], (n_samples - n0) * sizeof(float));
-
-            m_audio_pos = (m_audio_pos + n_samples) % m_audio.size();
-            m_audio_len = m_audio.size();
-        } else {
-            memcpy(&m_audio[m_audio_pos], stream, n_samples * sizeof(float));
-
-            m_audio_pos = (m_audio_pos + n_samples) % m_audio.size();
-            m_audio_len = std::min(m_audio_len + n_samples, m_audio.size());
-        }
-    }
-}
-
-void audio_async::get(int ms, std::vector<float> & result) {
-    if (!m_dev_id_in) {
-        fprintf(stderr, "%s: no audio device to get audio from!\n", __func__);
-        return;
-    }
-
-    if (!m_running) {
-        fprintf(stderr, "%s: not running!\n", __func__);
-        return;
-    }
-
-    result.clear();
-
-    {
-        std::lock_guard<std::mutex> lock(m_mutex);
-
-        if (ms <= 0) {
-            ms = m_len_ms;
-        }
-
-        size_t n_samples = (m_sample_rate * ms) / 1000;
-        if (n_samples > m_audio_len) {
-            n_samples = m_audio_len;
-        }
-
-        result.resize(n_samples);
-
-        int s0 = m_audio_pos - n_samples;
-        if (s0 < 0) {
-            s0 += m_audio.size();
-        }
-
-        if (s0 + n_samples > m_audio.size()) {
-            const size_t n0 = m_audio.size() - s0;
-
-            memcpy(result.data(), &m_audio[s0], n0 * sizeof(float));
-            memcpy(&result[n0], &m_audio[0], (n_samples - n0) * sizeof(float));
-        } else {
-            memcpy(result.data(), &m_audio[s0], n_samples * sizeof(float));
-        }
-    }
-}
-
-///////////////////////////
-
-std::string trim(const std::string & s) {
-    std::regex e("^\\s+|\\s+$");
-    return std::regex_replace(s, e, "");
-}
-
-std::string replace(const std::string & s, const std::string & from, const std::string & to) {
-    std::string result = s;
-    size_t pos = 0;
-    while ((pos = result.find(from, pos)) != std::string::npos) {
-        result.replace(pos, from.length(), to);
-        pos += to.length();
-    }
-    return result;
-}
-
-void high_pass_filter(std::vector<float> & data, float cutoff, float sample_rate) {
-    const float rc = 1.0f / (2.0f * M_PI * cutoff);
-    const float dt = 1.0f / sample_rate;
-    const float alpha = dt / (rc + dt);
-
-    float y = data[0];
-
-    for (size_t i = 1; i < data.size(); i++) {
-        y = alpha * (y + data[i] - data[i - 1]);
-        data[i] = y;
-    }
-}
-
-bool vad_simple(std::vector<float> & pcmf32, int sample_rate, int last_ms, float vad_thold, float freq_thold, bool verbose) {
-    const int n_samples      = pcmf32.size();
-    const int n_samples_last = (sample_rate * last_ms) / 1000;
-
-    if (n_samples_last >= n_samples) {
-        // not enough samples - assume no speech
-        return false;
-    }
-
-    if (freq_thold > 0.0f) {
-        high_pass_filter(pcmf32, freq_thold, sample_rate);
-    }
-
-    float energy_all  = 0.0f;
-    float energy_last = 0.0f;
-
-    for (int i = 0; i < n_samples; i++) {
-        energy_all += fabsf(pcmf32[i]);
-
-        if (i >= n_samples - n_samples_last) {
-            energy_last += fabsf(pcmf32[i]);
-        }
-    }
-
-    energy_all  /= n_samples;
-    energy_last /= n_samples_last;
-
-    if (verbose) {
-        fprintf(stderr, "%s: energy_all: %f, energy_last: %f, vad_thold: %f, freq_thold: %f\n", __func__, energy_all, energy_last, vad_thold, freq_thold);
-    }
-
-    if (energy_last > vad_thold*energy_all) {
-        return false;
-    }
-
-    return true;
-}
-
 std::string transcribe(whisper_context * ctx, const whisper_params & params, const std::vector<float> & pcmf32, float & prob, int64_t & t_ms) {
     const auto t_start = std::chrono::high_resolution_clock::now();
 
@@ -557,22 +241,10 @@ int main(int argc, char ** argv) {
     // main loop
     while (is_running) {
         // handle Ctrl + C
-        {
-            SDL_Event event;
-            while (SDL_PollEvent(&event)) {
-                switch (event.type) {
-                    case SDL_QUIT:
-                        {
-                            is_running = false;
-                        } break;
-                    default:
-                        break;
-                }
-            }
+        is_running = sdl_poll_events();
 
-            if (!is_running) {
-                break;
-            }
+        if (!is_running) {
+            break;
         }
 
         // delay
@@ -583,7 +255,7 @@ int main(int argc, char ** argv) {
         {
             audio.get(2000, pcmf32_cur);
 
-            if (vad_simple(pcmf32_cur, WHISPER_SAMPLE_RATE, 1250, params.vad_thold, params.freq_thold, params.print_energy) || force_speak) {
+            if (::vad_simple(pcmf32_cur, WHISPER_SAMPLE_RATE, 1250, params.vad_thold, params.freq_thold, params.print_energy) || force_speak) {
                 fprintf(stdout, "%s: Speech detected! Processing ...\n", __func__);
 
                 audio.get(params.voice_ms, pcmf32_cur);

examples/whisper.android/README.md

@@ -9,4 +9,4 @@ To use:
 5. Select the "release" active build variant, and use Android Studio to run and deploy to your device.
 [^1]: I recommend the tiny or base models for running on an Android device.
 
-<img width="300" alt="image" src="https://user-images.githubusercontent.com/1991296/208154256-82d972dc-221b-48c4-bfcb-36ce68602f93.png">
+<img width="300" alt="image" src="https://user-images.githubusercontent.com/1670775/221613663-a17bf770-27ef-45ab-9a46-a5f99ba65d2a.jpg">

examples/whisper.android/app/src/main/java/com/whispercppdemo/ui/main/MainScreen.kt

@@ -2,6 +2,7 @@ package com.whispercppdemo.ui.main
 
 import androidx.compose.foundation.layout.*
 import androidx.compose.foundation.rememberScrollState
+import androidx.compose.foundation.text.selection.SelectionContainer
 import androidx.compose.foundation.verticalScroll
 import androidx.compose.material3.*
 import androidx.compose.runtime.Composable
@@ -19,6 +20,7 @@ fun MainScreen(viewModel: MainScreenViewModel) {
         canTranscribe = viewModel.canTranscribe,
         isRecording = viewModel.isRecording,
         messageLog = viewModel.dataLog,
+        onBenchmarkTapped = viewModel::benchmark,
         onTranscribeSampleTapped = viewModel::transcribeSample,
         onRecordTapped = viewModel::toggleRecord
     )
@@ -30,6 +32,7 @@ private fun MainScreen(
     canTranscribe: Boolean,
     isRecording: Boolean,
     messageLog: String,
+    onBenchmarkTapped: () -> Unit,
     onTranscribeSampleTapped: () -> Unit,
     onRecordTapped: () -> Unit
 ) {
@@ -45,8 +48,11 @@ private fun MainScreen(
                 .padding(innerPadding)
                 .padding(16.dp)
         ) {
-            Row(horizontalArrangement = Arrangement.SpaceBetween) {
-                TranscribeSampleButton(enabled = canTranscribe, onClick = onTranscribeSampleTapped)
+            Column(verticalArrangement = Arrangement.SpaceBetween) {
+                Row(horizontalArrangement = Arrangement.SpaceBetween, modifier = Modifier.fillMaxWidth()) {
+                    BenchmarkButton(enabled = canTranscribe, onClick = onBenchmarkTapped)
+                    TranscribeSampleButton(enabled = canTranscribe, onClick = onTranscribeSampleTapped)
+                }
                 RecordButton(
                     enabled = canTranscribe,
                     isRecording = isRecording,
@@ -60,7 +66,16 @@ private fun MainScreen(
 
 @Composable
 private fun MessageLog(log: String) {
-    Text(modifier = Modifier.verticalScroll(rememberScrollState()), text = log)
+    SelectionContainer() {
+        Text(modifier = Modifier.verticalScroll(rememberScrollState()), text = log)
+    }
+}
+
+@Composable
+private fun BenchmarkButton(enabled: Boolean, onClick: () -> Unit) {
+    Button(onClick = onClick, enabled = enabled) {
+        Text("Benchmark")
+    }
 }
 
 @Composable

examples/whisper.android/app/src/main/java/com/whispercppdemo/ui/main/MainScreenViewModel.kt

@@ -41,10 +41,15 @@ class MainScreenViewModel(private val application: Application) : ViewModel() {
 
     init {
         viewModelScope.launch {
+            printSystemInfo()
             loadData()
         }
     }
 
+    private suspend fun printSystemInfo() {
+        printMessage(String.format("System Info: %s\n", WhisperContext.getSystemInfo()));
+    }
+
     private suspend fun loadData() {
         printMessage("Loading data...\n")
         try {
@@ -73,8 +78,7 @@ class MainScreenViewModel(private val application: Application) : ViewModel() {
         printMessage("Loading model...\n")
         val models = application.assets.list("models/")
         if (models != null) {
-            val inputstream = application.assets.open("models/" + models[0])
-            whisperContext = WhisperContext.createContextFromInputStream(inputstream)
+            whisperContext = WhisperContext.createContextFromAsset(application.assets, "models/" + models[0])
             printMessage("Loaded model ${models[0]}.\n")
         }
 
@@ -82,10 +86,29 @@ class MainScreenViewModel(private val application: Application) : ViewModel() {
         //whisperContext = WhisperContext.createContextFromFile(firstModel.absolutePath)
     }
 
+    fun benchmark() = viewModelScope.launch {
+        runBenchmark(6)
+    }
+
     fun transcribeSample() = viewModelScope.launch {
         transcribeAudio(getFirstSample())
     }
 
+    private suspend fun runBenchmark(nthreads: Int) {
+        if (!canTranscribe) {
+            return
+        }
+
+        canTranscribe = false
+
+        printMessage("Running benchmark. This will take minutes...\n")
+        whisperContext?.benchMemory(nthreads)?.let{ printMessage(it) }
+        printMessage("\n")
+        whisperContext?.benchGgmlMulMat(nthreads)?.let{ printMessage(it) }
+
+        canTranscribe = true
+    }
+
     private suspend fun getFirstSample(): File = withContext(Dispatchers.IO) {
         samplesPath.listFiles()!!.first()
     }
@@ -115,11 +138,14 @@ class MainScreenViewModel(private val application: Application) : ViewModel() {
         canTranscribe = false
 
         try {
-            printMessage("Reading wave samples...\n")
+            printMessage("Reading wave samples... ")
             val data = readAudioSamples(file)
+            printMessage("${data.size / (16000 / 1000)} ms\n")
             printMessage("Transcribing data...\n")
+            val start = System.currentTimeMillis()
             val text = whisperContext?.transcribeData(data)
-            printMessage("Done: $text\n")
+            val elapsed = System.currentTimeMillis() - start
+            printMessage("Done ($elapsed ms): $text\n")
         } catch (e: Exception) {
             Log.w(LOG_TAG, e)
             printMessage("${e.localizedMessage}\n")

examples/whisper.android/app/src/main/java/com/whispercppdemo/whisper/LibWhisper.kt

@@ -1,5 +1,6 @@
 package com.whispercppdemo.whisper
 
+import android.content.res.AssetManager
 import android.os.Build
 import android.util.Log
 import kotlinx.coroutines.*
@@ -26,6 +27,14 @@ class WhisperContext private constructor(private var ptr: Long) {
         }
     }
 
+    suspend fun benchMemory(nthreads: Int): String = withContext(scope.coroutineContext) {
+        return@withContext WhisperLib.benchMemcpy(nthreads)
+    }
+
+    suspend fun benchGgmlMulMat(nthreads: Int): String = withContext(scope.coroutineContext) {
+        return@withContext WhisperLib.benchGgmlMulMat(nthreads)
+    }
+
     suspend fun release() = withContext(scope.coroutineContext) {
         if (ptr != 0L) {
             WhisperLib.freeContext(ptr)
@@ -56,6 +65,19 @@ class WhisperContext private constructor(private var ptr: Long) {
             }
             return WhisperContext(ptr)
         }
+
+        fun createContextFromAsset(assetManager: AssetManager, assetPath: String): WhisperContext {
+            val ptr = WhisperLib.initContextFromAsset(assetManager, assetPath)
+
+            if (ptr == 0L) {
+                throw java.lang.RuntimeException("Couldn't create context from asset $assetPath")
+            }
+            return WhisperContext(ptr)
+        }
+
+        fun getSystemInfo(): String {
+            return WhisperLib.getSystemInfo()
+        }
     }
 }
 
@@ -64,6 +86,7 @@ private class WhisperLib {
         init {
             Log.d(LOG_TAG, "Primary ABI: ${Build.SUPPORTED_ABIS[0]}")
             var loadVfpv4 = false
+            var loadV8fp16 = false
             if (isArmEabiV7a()) {
                 // armeabi-v7a needs runtime detection support
                 val cpuInfo = cpuInfo()
@@ -74,11 +97,24 @@ private class WhisperLib {
                         loadVfpv4 = true
                     }
                 }
+            } else if (isArmEabiV8a()) {
+                // ARMv8.2a needs runtime detection support
+                val cpuInfo = cpuInfo()
+                cpuInfo?.let {
+                    Log.d(LOG_TAG, "CPU info: $cpuInfo")
+                    if (cpuInfo.contains("fphp")) {
+                        Log.d(LOG_TAG, "CPU supports fp16 arithmetic")
+                        loadV8fp16 = true
+                    }
+                }
             }
 
             if (loadVfpv4) {
                 Log.d(LOG_TAG, "Loading libwhisper_vfpv4.so")
                 System.loadLibrary("whisper_vfpv4")
+            } else if (loadV8fp16) {
+                Log.d(LOG_TAG, "Loading libwhisper_v8fp16_va.so")
+                System.loadLibrary("whisper_v8fp16_va")
             } else {
                 Log.d(LOG_TAG, "Loading libwhisper.so")
                 System.loadLibrary("whisper")
@@ -87,11 +123,15 @@ private class WhisperLib {
 
         // JNI methods
         external fun initContextFromInputStream(inputStream: InputStream): Long
+        external fun initContextFromAsset(assetManager: AssetManager, assetPath: String): Long
         external fun initContext(modelPath: String): Long
         external fun freeContext(contextPtr: Long)
         external fun fullTranscribe(contextPtr: Long, audioData: FloatArray)
         external fun getTextSegmentCount(contextPtr: Long): Int
         external fun getTextSegment(contextPtr: Long, index: Int): String
+        external fun getSystemInfo(): String
+        external fun benchMemcpy(nthread: Int): String
+        external fun benchGgmlMulMat(nthread: Int): String
     }
 }
 
@@ -99,6 +139,10 @@ private fun isArmEabiV7a(): Boolean {
     return Build.SUPPORTED_ABIS[0].equals("armeabi-v7a")
 }
 
+private fun isArmEabiV8a(): Boolean {
+    return Build.SUPPORTED_ABIS[0].equals("arm64-v8a")
+}
+
 private fun cpuInfo(): String? {
     return try {
         File("/proc/cpuinfo").inputStream().bufferedReader().use {

examples/whisper.android/app/src/main/jni/whisper/Android.mk

@@ -12,4 +12,15 @@ ifeq ($(TARGET_ARCH_ABI),armeabi-v7a)
 	# https://android.googlesource.com/platform/ndk/+/master/sources/android/cpufeatures/cpu-features.h
 	LOCAL_CFLAGS += -mfpu=neon-vfpv4
 	include $(BUILD_SHARED_LIBRARY)
-endif
+endif
+
+ifeq ($(TARGET_ARCH_ABI),arm64-v8a)
+	include $(CLEAR_VARS)
+	LOCAL_MODULE    := libwhisper_v8fp16_va
+	include $(LOCAL_PATH)/Whisper.mk
+	# Allow building NEON FMA code.
+	# https://android.googlesource.com/platform/ndk/+/master/sources/android/cpufeatures/cpu-features.h
+	LOCAL_CFLAGS += -march=armv8.2-a+fp16
+	include $(BUILD_SHARED_LIBRARY)
+endif
+

examples/whisper.android/app/src/main/jni/whisper/Whisper.mk

@@ -1,5 +1,5 @@
 WHISPER_LIB_DIR := $(LOCAL_PATH)/../../../../../../../
-LOCAL_LDLIBS    := -llog
+LOCAL_LDLIBS    := -landroid -llog
 
 # Make the final output library smaller by only keeping the symbols referenced from the app.
 ifneq ($(APP_OPTIM),debug)

examples/whisper.android/app/src/main/jni/whisper/jni.c

@@ -1,14 +1,18 @@
 #include <jni.h>
+#include <android/asset_manager.h>
+#include <android/asset_manager_jni.h>
 #include <android/log.h>
 #include <stdlib.h>
 #include <sys/sysinfo.h>
 #include <string.h>
 #include "whisper.h"
+#include "ggml.h"
 
 #define UNUSED(x) (void)(x)
 #define TAG "JNI"
 
 #define LOGI(...) __android_log_print(ANDROID_LOG_INFO,     TAG, __VA_ARGS__)
+#define LOGW(...) __android_log_print(ANDROID_LOG_WARN,     TAG, __VA_ARGS__)
 
 static inline int min(int a, int b) {
     return (a < b) ? a : b;
@@ -91,6 +95,52 @@ Java_com_whispercppdemo_whisper_WhisperLib_00024Companion_initContextFromInputSt
     return (jlong) context;
 }
 
+static size_t asset_read(void *ctx, void *output, size_t read_size) {
+    return AAsset_read((AAsset *) ctx, output, read_size);
+}
+
+static bool asset_is_eof(void *ctx) {
+    return AAsset_getRemainingLength64((AAsset *) ctx) <= 0;
+}
+
+static void asset_close(void *ctx) {
+    AAsset_close((AAsset *) ctx);
+}
+
+static struct whisper_context *whisper_init_from_asset(
+        JNIEnv *env,
+        jobject assetManager,
+        const char *asset_path
+) {
+    LOGI("Loading model from asset '%s'\n", asset_path);
+    AAssetManager *asset_manager = AAssetManager_fromJava(env, assetManager);
+    AAsset *asset = AAssetManager_open(asset_manager, asset_path, AASSET_MODE_STREAMING);
+    if (!asset) {
+        LOGW("Failed to open '%s'\n", asset_path);
+        return NULL;
+    }
+
+    whisper_model_loader loader = {
+            .context = asset,
+            .read = &asset_read,
+            .eof = &asset_is_eof,
+            .close = &asset_close
+    };
+
+    return whisper_init(&loader);
+}
+
+JNIEXPORT jlong JNICALL
+Java_com_whispercppdemo_whisper_WhisperLib_00024Companion_initContextFromAsset(
+        JNIEnv *env, jobject thiz, jobject assetManager, jstring asset_path_str) {
+    UNUSED(thiz);
+    struct whisper_context *context = NULL;
+    const char *asset_path_chars = (*env)->GetStringUTFChars(env, asset_path_str, NULL);
+    context = whisper_init_from_asset(env, assetManager, asset_path_chars);
+    (*env)->ReleaseStringUTFChars(env, asset_path_str, asset_path_chars);
+    return (jlong) context;
+}
+
 JNIEXPORT jlong JNICALL
 Java_com_whispercppdemo_whisper_WhisperLib_00024Companion_initContext(
         JNIEnv *env, jobject thiz, jstring model_path_str) {
@@ -164,4 +214,30 @@ Java_com_whispercppdemo_whisper_WhisperLib_00024Companion_getTextSegment(
     const char *text = whisper_full_get_segment_text(context, index);
     jstring string = (*env)->NewStringUTF(env, text);
     return string;
-}
+}
+
+JNIEXPORT jstring JNICALL
+Java_com_whispercppdemo_whisper_WhisperLib_00024Companion_getSystemInfo(
+        JNIEnv *env, jobject thiz
+) {
+    UNUSED(thiz);
+    const char *sysinfo = whisper_print_system_info();
+    jstring string = (*env)->NewStringUTF(env, sysinfo);
+    return string;
+}
+
+JNIEXPORT jstring JNICALL
+Java_com_whispercppdemo_whisper_WhisperLib_00024Companion_benchMemcpy(JNIEnv *env, jobject thiz,
+                                                                      jint n_threads) {
+    UNUSED(thiz);
+    const char *bench_ggml_memcpy = whisper_bench_memcpy_str(n_threads);
+    jstring string = (*env)->NewStringUTF(env, bench_ggml_memcpy);
+}
+
+JNIEXPORT jstring JNICALL
+Java_com_whispercppdemo_whisper_WhisperLib_00024Companion_benchGgmlMulMat(JNIEnv *env, jobject thiz,
+                                                                          jint n_threads) {
+    UNUSED(thiz);
+    const char *bench_ggml_mul_mat = whisper_bench_ggml_mul_mat_str(n_threads);
+    jstring string = (*env)->NewStringUTF(env, bench_ggml_mul_mat);
+}

examples/whisper.objc/README.md

@@ -24,3 +24,5 @@ Also, don't forget to add the `-DGGML_USE_ACCELERATE` compiler flag in Build Pha
 This can significantly improve the performance of the transcription:
 
 <img width="1072" alt="image" src="https://user-images.githubusercontent.com/1991296/208511239-8d7cdbd1-aa48-41b5-becd-ca288d53cc07.png">
+
+In this project, it also added `-O3 -DNDEBUG` to `Other C Flags`, but adding flags to app proj is not ideal in real world (applies to all C/C++ files), consider splitting xcodeproj in workspace in your own project.

examples/whisper.objc/whisper.objc.xcodeproj/project.pbxproj

@@ -296,6 +296,10 @@
 				IPHONEOS_DEPLOYMENT_TARGET = 16.0;
 				MTL_ENABLE_DEBUG_INFO = NO;
 				MTL_FAST_MATH = YES;
+				OTHER_CFLAGS = (
+					"-O3",
+					"-DNDEBUG",
+				);
 				SDKROOT = iphoneos;
 				VALIDATE_PRODUCT = YES;
 			};

examples/whisper.swiftui/README.md

@@ -1,14 +1,18 @@
 A sample SwiftUI app using [whisper.cpp](https://github.com/ggerganov/whisper.cpp/) to do voice-to-text transcriptions.
 See also: [whisper.objc](https://github.com/ggerganov/whisper.cpp/tree/master/examples/whisper.objc).
 
-To use:
+**Usage**:
 
 1. Select a model from the [whisper.cpp repository](https://github.com/ggerganov/whisper.cpp/tree/master/models).[^1]
-2. Add the model to "whisper.swiftui.demo/Resources/models" via Xcode.
+2. Add the model to `whisper.swiftui.demo/Resources/models` **via Xcode**.
 3. Select a sample audio file (for example, [jfk.wav](https://github.com/ggerganov/whisper.cpp/raw/master/samples/jfk.wav)).
-4. Add the model to "whisper.swiftui.demo/Resources/samples" via Xcode.
-5. Select the "release" build configuration under "Run", then deploy and run to your device.
+4. Add the sample audio file to `whisper.swiftui.demo/Resources/samples` **via Xcode**.
+5. Select the "Release" [^2] build configuration under "Run", then deploy and run to your device.
+
+**Note:** Pay attention to the folder path: `whisper.swiftui.demo/Resources/models` is the appropriate directory to place resources whilst `whisper.swiftui.demo/Models` is related to actual code.
 
 [^1]: I recommend the tiny, base or small models for running on an iOS device.
 
+[^2]: The `Release` build can boost performance of transcription. In this project, it also added `-O3 -DNDEBUG` to `Other C Flags`, but adding flags to app proj is not ideal in real world (applies to all C/C++ files), consider splitting xcodeproj in workspace in your own project.
+
 ![image](https://user-images.githubusercontent.com/1991296/212539216-0aef65e4-f882-480a-8358-0f816838fd52.png)

examples/whisper.swiftui/whisper.swiftui.xcodeproj/project.pbxproj

@@ -430,6 +430,10 @@
 				LLVM_LTO = YES;
 				MACOSX_DEPLOYMENT_TARGET = 13.0;
 				MARKETING_VERSION = 1.0;
+				OTHER_CFLAGS = (
+					"-O3",
+					"-DNDEBUG",
+				);
 				PRODUCT_BUNDLE_IDENTIFIER = com.whispercppdemo.WhisperCppDemo;
 				PRODUCT_NAME = "$(TARGET_NAME)";
 				SDKROOT = auto;

examples/whisper.wasm/CMakeLists.txt

@@ -32,8 +32,8 @@ set_target_properties(${TARGET} PROPERTIES LINK_FLAGS " \
     --bind \
     -s USE_PTHREADS=1 \
     -s PTHREAD_POOL_SIZE=8 \
-    -s INITIAL_MEMORY=1024MB \
-    -s TOTAL_MEMORY=1024MB \
+    -s INITIAL_MEMORY=1500MB \
+    -s TOTAL_MEMORY=1500MB \
     -s FORCE_FILESYSTEM=1 \
     -s EXPORTED_RUNTIME_METHODS=\"['print', 'printErr', 'ccall', 'cwrap']\" \
     ${EXTRA_FLAGS} \

examples/whisper.wasm/index-tmpl.html

@@ -46,10 +46,12 @@
 
             <div id="model">
                 Whisper model: <span id="model-whisper-status"></span>
-                <button id="fetch-whisper-tiny-en" onclick="loadWhisper('tiny.en')">tiny.en (75 MB)</button>
-                <button id="fetch-whisper-tiny"    onclick="loadWhisper('tiny')">tiny (75 MB)</button>
-                <button id="fetch-whisper-base-en" onclick="loadWhisper('base.en')">base.en (142 MB)</button>
-                <button id="fetch-whisper-base"    onclick="loadWhisper('base')">base (142 MB)</button>
+                <button id="fetch-whisper-tiny-en"  onclick="loadWhisper('tiny.en')">tiny.en (75 MB)</button>
+                <button id="fetch-whisper-tiny"     onclick="loadWhisper('tiny')">tiny (75 MB)</button>
+                <button id="fetch-whisper-base-en"  onclick="loadWhisper('base.en')">base.en (142 MB)</button>
+                <button id="fetch-whisper-base"     onclick="loadWhisper('base')">base (142 MB)</button>
+                <button id="fetch-whisper-small-en" onclick="loadWhisper('small.en')">small.en (466 MB)</button>
+                <button id="fetch-whisper-small"    onclick="loadWhisper('small')">small (466 MB)</button>
                 <span id="fetch-whisper-progress"></span>
 
                 <input type="file" id="whisper-file" name="file" onchange="loadFile(event, 'whisper.bin')" />
@@ -60,8 +62,8 @@
             <!-- radio button to select between file upload or microphone -->
             <div id="input">
                 Input:
-                <input type="radio" id="file" name="input" value="file" checked="checked" onchange="changeInput('file')" /> File
-                <input type="radio" id="mic" name="input" value="mic" onchange="changeInput('mic')" /> Microphone
+                <input type="radio" id="file" name="input" value="file" checked="checked" onchange="changeInput('file')" /> <label for="file">File</label>
+                <input type="radio" id="mic" name="input" value="mic" onchange="changeInput('mic')" /> <label for="mic">Microphone</label>
             </div>
 
             <br>
@@ -284,27 +286,33 @@
                 }
                 reader.readAsArrayBuffer(file);
 
-                document.getElementById('fetch-whisper-tiny-en').style.display = 'none';
-                document.getElementById('fetch-whisper-base-en').style.display = 'none';
-                document.getElementById('fetch-whisper-tiny'   ).style.display = 'none';
-                document.getElementById('fetch-whisper-base'   ).style.display = 'none';
-                document.getElementById('whisper-file'         ).style.display = 'none';
-                document.getElementById('model-whisper-status' ).innerHTML = 'loaded model: ' + file.name;
+                document.getElementById('fetch-whisper-tiny-en' ).style.display = 'none';
+                document.getElementById('fetch-whisper-base-en' ).style.display = 'none';
+                document.getElementById('fetch-whisper-small-en').style.display = 'none';
+                document.getElementById('fetch-whisper-tiny'    ).style.display = 'none';
+                document.getElementById('fetch-whisper-base'    ).style.display = 'none';
+                document.getElementById('fetch-whisper-small'   ).style.display = 'none';
+                document.getElementById('whisper-file'          ).style.display = 'none';
+                document.getElementById('model-whisper-status'  ).innerHTML = 'loaded model: ' + file.name;
             }
 
             function loadWhisper(model) {
                 let urls = {
-                    'tiny.en': 'https://whisper.ggerganov.com/ggml-model-whisper-tiny.en.bin',
-                    'tiny':    'https://whisper.ggerganov.com/ggml-model-whisper-tiny.bin',
-                    'base.en': 'https://whisper.ggerganov.com/ggml-model-whisper-base.en.bin',
-                    'base':    'https://whisper.ggerganov.com/ggml-model-whisper-base.bin',
+                    'tiny.en':  'https://whisper.ggerganov.com/ggml-model-whisper-tiny.en.bin',
+                    'tiny':     'https://whisper.ggerganov.com/ggml-model-whisper-tiny.bin',
+                    'base.en':  'https://whisper.ggerganov.com/ggml-model-whisper-base.en.bin',
+                    'base':     'https://whisper.ggerganov.com/ggml-model-whisper-base.bin',
+                    'small.en': 'https://whisper.ggerganov.com/ggml-model-whisper-small.en.bin',
+                    'small':    'https://whisper.ggerganov.com/ggml-model-whisper-small.bin',
                 };
 
                 let sizes = {
-                    'tiny.en': 75,
-                    'tiny':    75,
-                    'base.en': 142,
-                    'base':    142,
+                    'tiny.en':  75,
+                    'tiny':     75,
+                    'base.en':  142,
+                    'base':     142,
+                    'small.en': 466,
+                    'small':    466,
                 };
 
                 let url     = urls[model];
@@ -313,12 +321,14 @@
 
                 model_whisper = model;
 
-                document.getElementById('fetch-whisper-tiny-en').style.display = 'none';
-                document.getElementById('fetch-whisper-base-en').style.display = 'none';
-                document.getElementById('fetch-whisper-tiny'   ).style.display = 'none';
-                document.getElementById('fetch-whisper-base'   ).style.display = 'none';
-                document.getElementById('whisper-file'         ).style.display = 'none';
-                document.getElementById('model-whisper-status' ).innerHTML = 'loading model: ' + model;
+                document.getElementById('fetch-whisper-tiny-en' ).style.display = 'none';
+                document.getElementById('fetch-whisper-base-en' ).style.display = 'none';
+                document.getElementById('fetch-whisper-small-en').style.display = 'none';
+                document.getElementById('fetch-whisper-tiny'    ).style.display = 'none';
+                document.getElementById('fetch-whisper-base'    ).style.display = 'none';
+                document.getElementById('fetch-whisper-small'   ).style.display = 'none';
+                document.getElementById('whisper-file'          ).style.display = 'none';
+                document.getElementById('model-whisper-status'  ).innerHTML = 'loading model: ' + model;
 
                 cbProgress = function(p) {
                     let el = document.getElementById('fetch-whisper-progress');
@@ -327,12 +337,14 @@
 
                 cbCancel = function() {
                     var el;
-                    el = document.getElementById('fetch-whisper-tiny-en'); if (el) el.style.display = 'inline-block';
-                    el = document.getElementById('fetch-whisper-base-en'); if (el) el.style.display = 'inline-block';
-                    el = document.getElementById('fetch-whisper-tiny'   ); if (el) el.style.display = 'inline-block';
-                    el = document.getElementById('fetch-whisper-base'   ); if (el) el.style.display = 'inline-block';
-                    el = document.getElementById('whisper-file'         ); if (el) el.style.display = 'inline-block';
-                    el = document.getElementById('model-whisper-status' ); if (el) el.innerHTML = '';
+                    el = document.getElementById('fetch-whisper-tiny-en' ); if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('fetch-whisper-base-en' ); if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('fetch-whisper-small-en'); if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('fetch-whisper-tiny'    ); if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('fetch-whisper-base'    ); if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('fetch-whisper-small'   ); if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('whisper-file'          ); if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('model-whisper-status'  ); if (el) el.innerHTML = '';
                 };
 
                 loadRemote(url, dst, size_mb, cbProgress, storeFS, cbCancel, printTextarea);

examples/yt-wsp.sh

@@ -1,20 +1,10 @@
 #!/usr/bin/env bash
-
-# Small shell script to more easily automatically download and transcribe live stream VODs.
-# This uses YT-DLP, ffmpeg and the CPP version of Whisper: https://github.com/ggerganov/whisper.cpp
-# Use `./examples/yt-wsp.sh help` to print help info.
-#
-# Sample usage:
-#
-#   git clone https://github.com/ggerganov/whisper.cpp
-#   cd whisper.cpp
-#   make
-#   ./examples/yt-wsp.sh https://www.youtube.com/watch?v=1234567890
-#
+# shellcheck disable=2086
 
 # MIT License
 
 # Copyright (c) 2022 Daniils Petrovs
+# Copyright (c) 2023 Jennifer Capasso
 
 # Permission is hereby granted, free of charge, to any person obtaining a copy
 # of this software and associated documentation files (the "Software"), to deal
@@ -34,114 +24,178 @@
 # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 # SOFTWARE.
 
+# Small shell script to more easily automatically download and transcribe live stream VODs.
+# This uses YT-DLP, ffmpeg and the CPP version of Whisper: https://github.com/ggerganov/whisper.cpp
+# Use `./examples/yt-wsp.sh help` to print help info.
+#
+# Sample usage:
+#
+#   git clone https://github.com/ggerganov/whisper.cpp
+#   cd whisper.cpp
+#   make
+#   ./examples/yt-wsp.sh https://www.youtube.com/watch?v=1234567890
+#
+
 set -Eeuo pipefail
 
-# You can find how to download models in the OG repo: https://github.com/ggerganov/whisper.cpp/#usage
-MODEL_PATH="${MODEL_PATH:-models/ggml-base.en.bin}" # Set to a multilingual model if you want to translate from foreign lang to en
-WHISPER_EXECUTABLE="${WHISPER_EXECUTABLE:-whisper}" # Where to find the whisper.cpp executable
-WHISPER_LANG="${WHISPER_LANG:-en}" # Set to desired lang to translate from
+# get script file location
+SCRIPT_PATH="$(realpath -e ${BASH_SOURCE[0]})";
+SCRIPT_DIR="${SCRIPT_PATH%/*}"
+
+################################################################################
+# Documentation on downloading models can be found in the whisper.cpp repo:
+# https://github.com/ggerganov/whisper.cpp/#usage
+#
+# note: unless a multilingual model is specified, WHISPER_LANG will be ignored
+# and the video will be transcribed as if the audio were in the English language
+################################################################################
+MODEL_PATH="${MODEL_PATH:-${SCRIPT_DIR}/../models/ggml-base.en.bin}"
+
+################################################################################
+# Where to find the whisper.cpp executable.  default to the examples directory
+# which holds this script in source control
+################################################################################
+WHISPER_EXECUTABLE="${WHISPER_EXECUTABLE:-${SCRIPT_DIR}/../main}";
+
+# Set to desired language to be translated into english
+WHISPER_LANG="${WHISPER_LANG:-en}";
+
+# Default to 4 threads (this was most performant on my 2020 M1 MBP)
+WHISPER_THREAD_COUNT="${WHISPER_THREAD_COUNT:-4}";
 
 msg() {
     echo >&2 -e "${1-}"
 }
 
 cleanup() {
-    msg "Cleaning up..."
-    rm -rf "${temp_dir}" "vod-resampled.wav" "vod-resampled.wav.srt"
+    local -r clean_me="${1}";
+
+    if [ -d "${clean_me}" ]; then
+      msg "Cleaning up...";
+      rm -rf "${clean_me}";
+    else
+      msg "'${clean_me}' does not appear to be a directory!";
+      exit 1;
+    fi;
 }
 
 print_help() {
+    echo "################################################################################"
     echo "Usage: ./examples/yt-wsp.sh <video_url>"
-    echo "See configurable env variables in the script"
-    echo "This will produce an MP4 muxed file called res.mp4 in the working directory"
-    echo "Requirements: ffmpeg yt-dlp whisper"
-    echo "Whisper needs to be built into the main binary with make, then you can rename it to something like 'whisper' and add it to your PATH for convenience."
-    echo "E.g. in the root of Whisper.cpp, run: 'make && cp ./main /usr/local/bin/whisper'"
+    echo "# See configurable env variables in the script; there are many!"
+    echo "# This script will produce an MP4 muxed file in the working directory; it will"
+    echo "# be named for the title and id of the video."
+    echo "# passing in https://youtu.be/VYJtb2YXae8 produces a file named";
+    echo "# 'Why_we_all_need_subtitles_now-VYJtb2YXae8-res.mp4'"
+    echo "# Requirements: ffmpeg yt-dlp whisper.cpp"
+    echo "################################################################################"
 }
 
 check_requirements() {
     if ! command -v ffmpeg &>/dev/null; then
-        echo "ffmpeg is required (https://ffmpeg.org)."
+        echo "ffmpeg is required: https://ffmpeg.org";
         exit 1
-    fi
+    fi;
 
     if ! command -v yt-dlp &>/dev/null; then
-        echo "yt-dlp is required (https://github.com/yt-dlp/yt-dlp)."
-        exit 1
-    fi
+        echo "yt-dlp is required: https://github.com/yt-dlp/yt-dlp";
+        exit 1;
+    fi;
+
+    if ! command -v "${WHISPER_EXECUTABLE}" &>/dev/null; then
+        echo "The C++ implementation of Whisper is required: https://github.com/ggerganov/whisper.cpp"
+        echo "Sample usage:";
+        echo "";
+        echo "  git clone https://github.com/ggerganov/whisper.cpp";
+        echo "  cd whisper.cpp";
+        echo "  make";
+        echo "  ./examples/yt-wsp.sh https://www.youtube.com/watch?v=1234567890";
+        echo "";
+        exit 1;
+    fi;
 
-    if ! command -v "$WHISPER_EXECUTABLE" &>/dev/null; then
-        WHISPER_EXECUTABLE="./main"
-        if ! command -v "$WHISPER_EXECUTABLE" &>/dev/null; then
-            echo "Whisper is required (https://github.com/ggerganov/whisper.cpp):"
-            echo "Sample usage:"
-            echo ""
-            echo "  git clone https://github.com/ggerganov/whisper.cpp"
-            echo "  cd whisper.cpp"
-            echo "  make"
-            echo "  ./examples/yt-wsp.sh https://www.youtube.com/watch?v=1234567890"
-            echo ""
-            exit 1
-        fi
-    fi
 }
 
-if [[ $# -lt 1 ]]; then
-    print_help
-    exit 1
+if [[ "${#}" -lt 1 ]]; then
+    print_help;
+    exit 1;
 fi
 
-if [[ "$1" == "help" ]]; then
-    print_help
-    exit 0
+if [[ "${1##-*}" == "help" ]]; then
+    print_help;
+    exit 0;
 fi
 
-temp_dir="tmp"
-source_url="$1"
+check_requirements;
 
-check_requirements
+################################################################################
+# create a temporary directory to work in
+# set the temp_dir and temp_filename variables
+################################################################################
+temp_dir="$(mktemp -d ${SCRIPT_DIR}/tmp.XXXXXX)";
+temp_filename="${temp_dir}/yt-dlp-filename";
 
-msg "Downloading VOD..."
+################################################################################
+# for now we only take one argument
+# TODO: a for loop
+################################################################################
+source_url="${1}"
+title_name="";
 
-# Optionally add --cookies-from-browser BROWSER[+KEYRING][:PROFILE][::CONTAINER] for members only VODs
+msg "Downloading VOD...";
+
+################################################################################
+# Download the video, put the dynamic output filename into a variable.
+# Optionally add --cookies-from-browser BROWSER[+KEYRING][:PROFILE][::CONTAINER]
+# for videos only available to logged-in users.
+################################################################################
 yt-dlp \
     -f "bestvideo[ext=mp4]+bestaudio[ext=m4a]/best[ext=mp4]/best" \
+    -o "${temp_dir}/%(title)s-%(id)s.vod.mp4" \
+    --print-to-file "%(filename)s" "${temp_filename}" \
+    --no-simulate \
+    --no-write-auto-subs \
+    --restrict-filenames \
     --embed-thumbnail \
     --embed-chapters \
     --xattrs \
-    "${source_url}" -o "${temp_dir}/vod.mp4"
+    "${source_url}";
 
-msg "Extracting audio and resampling..."
+title_name="$(xargs basename -s .vod.mp4 < ${temp_filename})";
 
-ffmpeg -i "${temp_dir}/vod.mp4" \
+msg "Extracting audio and resampling...";
+
+ffmpeg -i "${temp_dir}/${title_name}.vod.mp4"  \
     -hide_banner \
+    -vn \
     -loglevel error \
     -ar 16000 \
     -ac 1 \
-    -c:a \
-    pcm_s16le -y "vod-resampled.wav"
+    -c:a pcm_s16le \
+    -y \
+    "${temp_dir}/${title_name}.vod-resampled.wav";
 
-msg "Transcribing to subtitle file..."
-msg "Whisper specified at: ${WHISPER_EXECUTABLE}"
+msg "Transcribing to subtitle file...";
+msg "Whisper specified at: '${WHISPER_EXECUTABLE}'";
 
-$WHISPER_EXECUTABLE \
+"${WHISPER_EXECUTABLE}" \
     -m "${MODEL_PATH}" \
     -l "${WHISPER_LANG}" \
-    -f "vod-resampled.wav" \
-    -t 8 \
+    -f "${temp_dir}/${title_name}.vod-resampled.wav" \
+    -t "${WHISPER_THREAD_COUNT}" \
     -osrt \
-    --translate
+    --translate;
 
-msg "Embedding subtitle track..."
+msg "Embedding subtitle track...";
 
-ffmpeg -i "${temp_dir}/vod.mp4" \
+ffmpeg -i "${temp_dir}/${title_name}.vod.mp4" \
     -hide_banner \
     -loglevel error \
-    -i "vod-resampled.wav.srt" \
+    -i "${temp_dir}/${title_name}.vod-resampled.wav.srt" \
     -c copy \
     -c:s mov_text \
-    -y res.mp4
+    -y "${title_name}-res.mp4";
 
-cleanup
+#cleanup "${temp_dir}";
 
-msg "Done! Your finished file is ready: res.mp4"
+msg "Done! Your finished file is ready: ${title_name}-res.mp4";

extra/bench-all.sh

@@ -12,6 +12,18 @@ fi
 
 models=( "tiny" "base" "small" "medium" "large" )
 
+printf "\n"
+printf "Running memcpy benchmark with 1 thread\n"
+printf "\n"
+
+./bench -w 1 -t 1 2>&1
+
+printf "\n"
+printf "Running ggml_mul_mat benchmark with $n_threads threads\n"
+printf "\n"
+
+./bench -w 2 -t $n_threads 2>&1
+
 printf "\n"
 printf "Running benchmark for all models\n"
 printf "This can take a while!\n"
@@ -52,8 +64,11 @@ for model in "${models[@]}"; do
         config="$config BLAS"
     fi
 
+    if [[ $system_info == *"COREML = 1"* ]]; then
+        config="$config COREML"
+    fi
+
     commit=$(git rev-parse --short HEAD)
 
     printf "| <todo> | <todo> | $config | $model | $n_threads | $load_time | $encode_time | $commit |\n"
 done
-

extra/bench-wts.sh

@@ -0,0 +1,70 @@
+# Benchmark word-level timestamps for different models
+#
+# This script takes two arguments
+# - an audio file
+# - [optional] path to a font file
+
+# I'm using "/usr/share/fonts/truetype/freefont/FreeMono.ttf" on Ubuntu
+
+if [ -z "$1" ]; then
+    echo "Usage: $0 <audio file> [font file]"
+    exit 1
+fi
+
+#TODO: Make this a command line parameter
+#models="base small large"
+#models="tiny.en tiny base.en base small.en small medium.en medium large-v1 large"
+models="tiny.en base.en small.en medium.en large"
+
+DURATION=$(ffprobe -i $1 -show_entries format=duration -v quiet -of csv="p=0")
+DURATION=$(printf "%.2f" $DURATION)
+echo "Input file duration: ${DURATION}s"
+
+for model in $models; do
+    echo "Running $model"
+    COMMAND="./main -m models/ggml-$model.bin -owts -f $1 -of $1.$model"
+
+    if [ ! -z "$2" ]; then
+        COMMAND="$COMMAND -fp $2"
+    fi
+    #TODO: Surface errors better
+    # TIMEFMT is for zsh, TIMEFORMAT is for bash
+    EXECTIME=$({ TIMEFMT="%E";TIMEFORMAT=%E; time $COMMAND >/dev/null 2>&1; } 2>&1)
+
+    # Slightly different formats between zsh and bash
+    if [ "${EXECTIME: -1}" == "s" ]; then
+        EXECTIME=${EXECTIME::-1}
+    fi
+
+    RATIO=$(echo "$DURATION / $EXECTIME" | bc -l)
+    RATIO=$(printf "%.2f" $RATIO)
+
+    echo "Execution time: ${EXECTIME}s (${RATIO}x realtime)"
+
+    # If the file already exists, delete it
+    if [ -f $1.mp4 ]; then
+        rm $1.mp4
+    fi
+
+    bash $1.$model.wts >/dev/null 2>&1
+    mv $1.mp4 $1.$model.mp4
+
+    ffmpeg -y -f lavfi -i color=c=black:s=1200x50:d=$DURATION -vf "drawtext=fontfile=$2:fontsize=36:x=10:y=(h-text_h)/2:text='ggml-$model - ${EXECTIME}s (${RATIO}x realtime)':fontcolor=lightgrey" $1.$model.info.mp4 >/dev/null 2>&1
+done
+
+COMMAND="ffmpeg -y"
+for model in $models; do
+    COMMAND="$COMMAND -i $1.$model.info.mp4 -i $1.$model.mp4"
+done
+COMMAND="$COMMAND -filter_complex \""
+COUNT=0
+for model in $models; do
+    COMMAND="$COMMAND[${COUNT}:v][$(($COUNT+1)):v]"
+    COUNT=$((COUNT+2))
+done
+COMMAND="$COMMAND vstack=inputs=${COUNT}[v]\" -map \"[v]\" -map 1:a $1.all.mp4 >/dev/null 2>&1"
+
+echo $COMMAND
+
+# Run the command
+eval $COMMAND

ggml.h

@@ -177,11 +177,12 @@ extern "C" {
 #include <stddef.h>
 #include <stdbool.h>
 
-#define GGML_MAX_DIMS     4
-#define GGML_MAX_NODES    4096
-#define GGML_MAX_PARAMS   16
-#define GGML_MAX_CONTEXTS 64
-#define GGML_MAX_OPT      4
+#define GGML_MAX_DIMS          4
+#define GGML_MAX_NODES         4096
+#define GGML_MAX_PARAMS        16
+#define GGML_MAX_CONTEXTS      64
+#define GGML_MAX_OPT           4
+#define GGML_DEFAULT_N_THREADS 4
 
 #ifdef __ARM_NEON
 // we use the built-in 16-bit float type
@@ -198,11 +199,14 @@ struct ggml_object;
 struct ggml_context;
 
 enum ggml_type {
+    // explicitly numbered values are used in llama.cpp files
+    GGML_TYPE_F32  = 0,
+    GGML_TYPE_F16  = 1,
+    GGML_TYPE_Q4_0 = 2,
+    GGML_TYPE_Q4_1 = 3,
     GGML_TYPE_I8,
     GGML_TYPE_I16,
     GGML_TYPE_I32,
-    GGML_TYPE_F16,
-    GGML_TYPE_F32,
     GGML_TYPE_COUNT,
 };
 
@@ -226,12 +230,15 @@ enum ggml_op {
     GGML_OP_STEP,
     GGML_OP_RELU,
     GGML_OP_GELU,
+    GGML_OP_SILU,
     GGML_OP_NORM, // normalize
+    GGML_OP_RMS_NORM,
 
     GGML_OP_MUL_MAT,
 
     GGML_OP_SCALE,
     GGML_OP_CPY,
+    GGML_OP_CONT,
     GGML_OP_RESHAPE,
     GGML_OP_VIEW,
     GGML_OP_PERMUTE,
@@ -246,19 +253,35 @@ enum ggml_op {
     GGML_OP_FLASH_ATTN,
     GGML_OP_FLASH_FF,
 
+    GGML_OP_MAP_UNARY,
+    GGML_OP_MAP_BINARY,
+
     GGML_OP_COUNT,
 };
 
+
+// ggml object
+struct ggml_object {
+    size_t offs;
+    size_t size;
+
+    struct ggml_object * next;
+
+    char padding[8];
+};
+
+static const size_t GGML_OBJECT_SIZE = sizeof(struct ggml_object);
+
 // n-dimensional tensor
 struct ggml_tensor {
     enum ggml_type type;
 
     int    n_dims;
-    int    ne[GGML_MAX_DIMS]; // number of elements
-    size_t nb[GGML_MAX_DIMS]; // stride in bytes:
-                              // nb[0] = sizeof(type)
-                              // nb[1] = nb[0]   * ne[0] + padding
-                              // nb[i] = nb[i-1] * ne[i-1]
+    int64_t ne[GGML_MAX_DIMS]; // number of elements
+    size_t  nb[GGML_MAX_DIMS]; // stride in bytes:
+                               // nb[0] = sizeof(type)
+                               // nb[1] = nb[0]   * ne[0] + padding
+                               // nb[i] = nb[i-1] * ne[i-1]
 
     // compute data
     enum ggml_op op;
@@ -301,10 +324,18 @@ struct ggml_cgraph {
     int64_t perf_time_us;
 };
 
+// scratch buffer
+struct ggml_scratch {
+    size_t offs;
+    size_t size;
+    void * data;
+};
+
 struct ggml_init_params {
     // memory pool
     size_t mem_size;   // bytes
     void * mem_buffer; // if NULL, memory will be allocated internally
+    bool   no_alloc;   // don't allocate memory for the tensor data
 };
 
 void    ggml_time_init(void); // call this once at the beginning of the program
@@ -316,10 +347,13 @@ int64_t ggml_cycles_per_ms(void);
 void ggml_print_object (const struct ggml_object * obj);
 void ggml_print_objects(const struct ggml_context * ctx);
 
-int    ggml_nelements(const struct ggml_tensor * tensor);
-size_t ggml_nbytes   (const struct ggml_tensor * tensor);
+int64_t ggml_nelements(const struct ggml_tensor * tensor);
+size_t  ggml_nbytes   (const struct ggml_tensor * tensor);
+
+int    ggml_blck_size (enum ggml_type type);
+size_t ggml_type_size (enum ggml_type type); // size in bytes for all elements in a block
+float  ggml_type_sizef(enum ggml_type type); // ggml_type_size()/ggml_blck_size() as float
 
-size_t ggml_type_size   (enum ggml_type type);
 size_t ggml_element_size(const struct ggml_tensor * tensor);
 
 struct ggml_context * ggml_init(struct ggml_init_params params);
@@ -327,37 +361,39 @@ void ggml_free(struct ggml_context * ctx);
 
 size_t ggml_used_mem(const struct ggml_context * ctx);
 
+size_t ggml_set_scratch(struct ggml_context * ctx, struct ggml_scratch scratch);
+
 struct ggml_tensor * ggml_new_tensor(
         struct ggml_context * ctx,
         enum   ggml_type type,
         int    n_dims,
-        const int *ne);
+        const int64_t *ne);
 
 struct ggml_tensor * ggml_new_tensor_1d(
         struct ggml_context * ctx,
         enum   ggml_type type,
-        int    ne0);
+        int64_t ne0);
 
 struct ggml_tensor * ggml_new_tensor_2d(
         struct ggml_context * ctx,
         enum   ggml_type type,
-        int    ne0,
-        int    ne1);
+        int64_t ne0,
+        int64_t ne1);
 
 struct ggml_tensor * ggml_new_tensor_3d(
         struct ggml_context * ctx,
         enum   ggml_type type,
-        int    ne0,
-        int    ne1,
-        int    ne2);
+        int64_t ne0,
+        int64_t ne1,
+        int64_t ne2);
 
 struct ggml_tensor * ggml_new_tensor_4d(
         struct ggml_context * ctx,
         enum   ggml_type type,
-        int    ne0,
-        int    ne1,
-        int    ne2,
-        int    ne3);
+        int64_t ne0,
+        int64_t ne1,
+        int64_t ne2,
+        int64_t ne3);
 
 struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value);
 struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value);
@@ -457,12 +493,20 @@ struct ggml_tensor * ggml_gelu(
         struct ggml_context * ctx,
         struct ggml_tensor  * a);
 
+struct ggml_tensor * ggml_silu(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a);
+
 // normalize along rows
 // TODO: eps is hardcoded to 1e-5 for now
 struct ggml_tensor * ggml_norm(
         struct ggml_context * ctx,
         struct ggml_tensor  * a);
 
+struct ggml_tensor * ggml_rms_norm(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a);
+
 // A: m rows, n columns
 // B: p rows, n columns (i.e. we transpose it internally)
 // result is m columns, p rows
@@ -487,6 +531,11 @@ struct ggml_tensor * ggml_cpy(
         struct ggml_tensor  * a,
         struct ggml_tensor  * b);
 
+// make contiguous
+struct ggml_tensor * ggml_cont(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a);
+
 // return view(a), b specifies the new shape
 // TODO: when we start computing gradient, make a copy instead of view
 struct ggml_tensor * ggml_reshape(
@@ -499,33 +548,43 @@ struct ggml_tensor * ggml_reshape(
 struct ggml_tensor * ggml_reshape_2d(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        int                   ne0,
-        int                   ne1);
+        int64_t               ne0,
+        int64_t               ne1);
 
 // return view(a)
 // TODO: when we start computing gradient, make a copy instead of view
 struct ggml_tensor * ggml_reshape_3d(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        int                   ne0,
-        int                   ne1,
-        int                   ne2);
+        int64_t               ne0,
+        int64_t               ne1,
+        int64_t               ne2);
 
 // offset in bytes
 struct ggml_tensor * ggml_view_1d(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        int                   ne0,
+        int64_t               ne0,
         size_t                offset);
 
 struct ggml_tensor * ggml_view_2d(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        int                   ne0,
-        int                   ne1,
+        int64_t               ne0,
+        int64_t               ne1,
         size_t                nb1, // row stride in bytes
         size_t                offset);
 
+struct ggml_tensor * ggml_view_3d(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int64_t               ne0,
+        int64_t               ne1,
+        int64_t               ne2,
+        size_t                nb1, // row   stride in bytes
+        size_t                nb2, // slice stride in bytes
+        size_t                offset);
+
 struct ggml_tensor * ggml_permute(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
@@ -596,6 +655,21 @@ struct ggml_tensor * ggml_flash_ff(
         struct ggml_tensor  * c0,
         struct ggml_tensor  * c1);
 
+// Mapping operations
+typedef void (*ggml_unary_op_f32_t)(const int, float *, const float *);
+typedef void (*ggml_binary_op_f32_t)(const int, float *, const float *, const float *);
+
+struct ggml_tensor * ggml_map_unary_f32(
+        struct ggml_context        * ctx,
+        struct ggml_tensor         * a,
+        const  ggml_unary_op_f32_t fun);
+
+struct ggml_tensor * ggml_map_binary_f32(
+        struct ggml_context         * ctx,
+        struct ggml_tensor          * a,
+        struct ggml_tensor          * b,
+        const  ggml_binary_op_f32_t fun);
+
 //
 // automatic differentiation
 //
@@ -717,6 +791,13 @@ enum ggml_opt_result ggml_opt(
         struct ggml_opt_params params,
         struct ggml_tensor * f);
 
+//
+// quantization
+//
+
+size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int64_t * hist);
+size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * hist);
+
 //
 // system info
 //
@@ -734,6 +815,30 @@ int ggml_cpu_has_blas(void);
 int ggml_cpu_has_sse3(void);
 int ggml_cpu_has_vsx(void);
 
+
+//
+// Internal types and functions exposed for tests and benchmarks
+//
+
+#ifdef  __cplusplus
+// restrict not standard in C++
+#define GGML_RESTRICT
+#else
+#define GGML_RESTRICT restrict
+#endif
+typedef void (*dequantize_row_q_t)(const void * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
+typedef void (*quantize_row_q_t)(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
+typedef void (*vec_dot_q_t)(const int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT x, const void * GGML_RESTRICT y);
+
+typedef struct {
+    dequantize_row_q_t dequantize_row_q;
+    quantize_row_q_t   quantize_row_q;
+    quantize_row_q_t   quantize_row_q_reference;
+    vec_dot_q_t        vec_dot_q;
+} quantize_fns_t;
+
+quantize_fns_t ggml_internal_get_quantize_fn(size_t i);
+
 #ifdef  __cplusplus
 }
 #endif

models/README.md

@@ -6,7 +6,7 @@ using the [convert-pt-to-ggml.py](convert-pt-to-ggml.py) script. You can either
 the `ggml` files yourself using the conversion script, or you can use the [download-ggml-model.sh](download-ggml-model.sh)
 script to download the already converted models. Currently, they are hosted on the following locations:
 
-- https://huggingface.co/datasets/ggerganov/whisper.cpp
+- https://huggingface.co/ggerganov/whisper.cpp
 - https://ggml.ggerganov.com
 
 Sample usage:
@@ -23,7 +23,7 @@ You can now use it like this:
 
 A third option to obtain the model files is to download them from Hugging Face:
 
-https://huggingface.co/datasets/ggerganov/whisper.cpp/tree/main
+https://huggingface.co/ggerganov/whisper.cpp/tree/main
 
 ## Available models
 

models/convert-h5-to-ggml.py

@@ -79,11 +79,11 @@ dir_model   = sys.argv[1]
 dir_whisper = sys.argv[2]
 dir_out     = sys.argv[3]
 
-with open(dir_model + "/vocab.json", "r") as f:
+with open(dir_model + "/vocab.json", "r", encoding="utf8") as f:
     encoder = json.load(f)
-with open(dir_model + "/added_tokens.json", "r") as f:
+with open(dir_model + "/added_tokens.json", "r", encoding="utf8") as f:
     encoder_added = json.load(f)
-with open(dir_model + "/config.json", "r") as f:
+with open(dir_model + "/config.json", "r", encoding="utf8") as f:
     hparams = json.load(f)
 
 model = WhisperForConditionalGeneration.from_pretrained(dir_model)

models/convert-pt-to-ggml.py

@@ -39,6 +39,7 @@ import json
 import code
 import torch
 import numpy as np
+import base64
 
 #from transformers import GPTJForCausalLM
 #from transformers import GPT2TokenizerFast
@@ -224,18 +225,14 @@ with np.load(os.path.join(dir_whisper, "whisper/assets", "mel_filters.npz")) as
 #code.interact(local=locals())
 
 multilingual = hparams["n_vocab"] == 51865
-dir_tokenizer = os.path.join(dir_whisper, "whisper/assets", multilingual and "multilingual" or "gpt2")
-
-#tokenizer = build_tokenizer(dir_whisper, multilingual and "multilingual" or "gpt2")
-#print(tokenizer)
-#print(tokenizer.name_or_path)
-#print(len(tokenizer.additional_special_tokens))
+tokenizer = os.path.join(dir_whisper, "whisper/assets", multilingual and "multilingual.tiktoken" or "gpt2.tiktoken")
 
 # output in the same directory as the model
 fname_out = dir_out + "/ggml-model.bin"
 
-with open(dir_tokenizer + "/vocab.json", "r", encoding="utf8") as f:
-    tokens = json.load(f)
+with open(tokenizer, "rb") as f:
+    contents = f.read()
+    tokens = {base64.b64decode(token): int(rank) for token, rank in (line.split() for line in contents.splitlines() if line)}
 
 # use 16-bit or 32-bit floats
 use_f16 = True
@@ -271,9 +268,8 @@ byte_decoder = {v:k for k, v in byte_encoder.items()}
 fout.write(struct.pack("i", len(tokens)))
 
 for key in tokens:
-    text = bytearray([byte_decoder[c] for c in key])
-    fout.write(struct.pack("i", len(text)))
-    fout.write(text)
+    fout.write(struct.pack("i", len(key)))
+    fout.write(key)
 
 for name in list_vars.keys():
     data = list_vars[name].squeeze().numpy()

models/convert-whisper-to-coreml.py

@@ -0,0 +1,334 @@
+import argparse
+import torch
+import torch.nn.functional as F
+import coremltools as ct
+
+from torch import Tensor
+from torch import nn
+from typing import Dict
+from typing import Optional
+from ane_transformers.reference.layer_norm import LayerNormANE as LayerNormANEBase
+from coremltools.models.neural_network.quantization_utils import quantize_weights
+from whisper.model import Whisper, AudioEncoder, TextDecoder, ResidualAttentionBlock, MultiHeadAttention, ModelDimensions
+from whisper import load_model
+
+# Use for changing dim of input in encoder and decoder embeddings
+def linear_to_conv2d_map(state_dict, prefix, local_metadata, strict,
+                         missing_keys, unexpected_keys, error_msgs):
+    """
+    Unsqueeze twice to map nn.Linear weights to nn.Conv2d weights
+    """
+    for k in state_dict:
+        is_attention = all(substr in k for substr in ['attn', '.weight'])
+        is_mlp = any([k.endswith(s) for s in ['mlp.0.weight', 'mlp.2.weight']])
+
+        if (is_attention or is_mlp) and len(state_dict[k].shape) == 2:
+            state_dict[k] = state_dict[k][:, :, None, None]
+
+
+def correct_for_bias_scale_order_inversion(state_dict, prefix, local_metadata,
+                                           strict, missing_keys,
+                                           unexpected_keys, error_msgs):
+    state_dict[prefix + 'bias'] = state_dict[prefix + 'bias'] / state_dict[prefix + 'weight']
+    return state_dict
+
+class LayerNormANE(LayerNormANEBase):
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self._register_load_state_dict_pre_hook(
+            correct_for_bias_scale_order_inversion)
+
+class MultiHeadAttentionANE(MultiHeadAttention):
+    def __init__(self, n_state: int, n_head: int):
+        super().__init__(n_state, n_head)
+
+        setattr(self, 'query', nn.Conv2d(n_state, n_state, kernel_size=1))
+        setattr(self, 'key', nn.Conv2d(n_state, n_state, kernel_size=1, bias=False))
+        setattr(self, 'value', nn.Conv2d(n_state, n_state, kernel_size=1))
+        setattr(self, 'out', nn.Conv2d(n_state, n_state, kernel_size=1))
+
+    def forward(self,
+                x: Tensor,
+                xa: Optional[Tensor] = None,
+                mask: Optional[Tensor] = None,
+                kv_cache: Optional[dict] = None):
+
+        q = self.query(x)
+
+        if kv_cache is None or xa is None or self.key not in kv_cache:
+            # hooks, if installed (i.e. kv_cache is not None), will prepend the cached kv tensors;
+            # otherwise, perform key/value projections for self- or cross-attention as usual.
+            k = self.key(x if xa is None else xa)
+            v = self.value(x if xa is None else xa)
+
+        else:
+            # for cross-attention, calculate keys and values once and reuse in subsequent calls.
+            k = kv_cache[self.key]
+            v = kv_cache[self.value]
+
+        wv, qk = self.qkv_attention_ane(q, k, v, mask)
+
+        return self.out(wv), qk
+
+    def qkv_attention_ane(self, q: Tensor, k: Tensor, v: Tensor, mask: Optional[Tensor] = None):
+
+        _, dim, _, seqlen = q.size()
+
+        dim_per_head = dim // self.n_head
+
+        scale = float(dim_per_head)**-0.5
+
+        q = q * scale
+
+        mh_q = q.split(dim_per_head, dim=1)
+        mh_k = k.transpose(1,3).split(dim_per_head, dim=3)
+        mh_v = v.split(dim_per_head, dim=1)
+
+        mh_qk = [
+            torch.einsum('bchq,bkhc->bkhq', [qi, ki])
+            for qi, ki in zip(mh_q, mh_k)
+        ]  # (batch_size, max_seq_length, 1, max_seq_length) * n_heads
+
+        if mask is not None:
+            for head_idx in range(self.n_head):
+                mh_qk[head_idx] = mh_qk[head_idx] + mask[:, :seqlen, :, :seqlen]
+
+        attn_weights = [aw.softmax(dim=1) for aw in mh_qk]  # (batch_size, max_seq_length, 1, max_seq_length) * n_heads
+        attn = [torch.einsum('bkhq,bchk->bchq', wi, vi) for wi, vi in zip(attn_weights, mh_v)]  # (batch_size, dim_per_head, 1, max_seq_length) * n_heads
+        attn = torch.cat(attn, dim=1)  # (batch_size, dim, 1, max_seq_length)
+
+        return attn, torch.cat(mh_qk, dim=1).float().detach()
+
+
+class ResidualAttentionBlockANE(ResidualAttentionBlock):
+    def __init__(self, n_state: int, n_head: int, cross_attention: bool = False):
+        super().__init__(n_state, n_head, cross_attention)
+
+        setattr(self, 'attn', MultiHeadAttentionANE(n_state, n_head))
+        setattr(self, 'attn_ln', LayerNormANE(n_state))
+
+        setattr(self, 'cross_attn', MultiHeadAttentionANE(n_state, n_head) if cross_attention else None)
+        setattr(self, 'cross_attn_ln', LayerNormANE(n_state) if cross_attention else None)
+
+        n_mlp = n_state * 4
+        setattr(self, 'mlp', nn.Sequential(
+            nn.Conv2d(n_state, n_mlp, kernel_size=1),
+            nn.GELU(),
+            nn.Conv2d(n_mlp, n_state, kernel_size=1)
+        ))
+        setattr(self, 'mlp_ln', LayerNormANE(n_state))
+
+
+class AudioEncoderANE(AudioEncoder):
+    def __init__(self, n_mels: int, n_ctx: int, n_state: int, n_head: int, n_layer: int):
+        super().__init__(n_mels, n_ctx, n_state, n_head, n_layer)
+
+        setattr(self, 'blocks', nn.ModuleList(
+            [ResidualAttentionBlockANE(n_state, n_head) for _ in range(n_layer)]
+        ))
+        setattr(self, 'ln_post', LayerNormANE(n_state))
+
+    def forward(self, x: Tensor):
+        """
+        x : torch.Tensor, shape = (batch_size, n_mels, n_ctx)
+            the mel spectrogram of the audio
+        """
+        x = F.gelu(self.conv1(x))
+        x = F.gelu(self.conv2(x))
+
+        assert x.shape[1:] == self.positional_embedding.shape[::-1], "incorrect audio shape"
+
+        # Add positional embedding and add dummy dim for ANE
+        x = (x + self.positional_embedding.transpose(0,1)).to(x.dtype).unsqueeze(2)
+
+        for block in self.blocks:
+            x = block(x)
+
+        x = self.ln_post(x)
+
+        # """
+        # TODO:
+        # I think we need to transpose the result here to make it fit whisper.cpp memory order.
+        # However, even doing this, the results are still wrong. Kind of less wrong compared to
+        # not transposing, but still wrong.
+
+        # Also, I don't know why the original OpenAI implementation does not need to transpose
+
+        # transpose to (batch_size, n_ctx, n_state)
+        # x : torch.Tensor, shape = (batch_size, n_state, 1, n_ctx)
+
+        # """
+        # x = x.transpose(1,3)
+
+        return x
+
+class TextDecoderANE(TextDecoder):
+
+    def __init__(self, n_vocab: int, n_ctx: int, n_state: int, n_head: int, n_layer: int):
+        super().__init__(n_vocab, n_ctx, n_state, n_head, n_layer)
+
+        setattr(self, 'blocks', nn.ModuleList(
+            [ResidualAttentionBlockANE(n_state, n_head, cross_attention=True) for _ in range(n_layer)]
+        ))
+        setattr(self, 'ln', LayerNormANE(n_state))
+
+    def forward(self, x: Tensor, xa: Tensor, kv_cache: Optional[dict] = None):
+        """
+        x : torch.LongTensor, shape = (batch_size, <= n_ctx)
+            the text tokens
+        xa : torch.Tensor, shape = (batch_size, n_mels, n_audio_ctx)
+            the encoded audio features to be attended on
+        """
+        offset = next(iter(kv_cache.values())).shape[3] if kv_cache else 0
+        x = self.token_embedding(x) + self.positional_embedding[offset : offset + x.shape[-1]]
+        x = x.to(xa.dtype)
+
+        # Reformat for ANE
+        mask = self.mask[None, None, :, :].permute(0,3,1,2)
+        x = x.transpose(1,2).unsqueeze(2)
+
+        for block in self.blocks:
+            x = block(x, xa, mask=mask, kv_cache=kv_cache)
+
+        x = self.ln(x)
+
+        # Reformat back from ANE
+        x = x.permute(0,2,3,1).squeeze(0)
+
+        # ANE can only load tensors with dim size of at most 16,384 - whisper uses 51,864 (en) or 51,865 (multi-lang) tokens so we need to compute in chunks
+        if self.token_embedding.weight.shape[0] == 51865:
+            # split in 11 chunks - 4715 each
+            splits = self.token_embedding.weight.split(self.token_embedding.weight.shape[0]//11, dim=0)
+            logits = torch.cat([torch.einsum('bid,jd->bij', x, split) for split in splits]).view(*x.shape[:2], -1)
+        else:
+            # split in 12 chunks - 4322 each
+            assert(self.token_embedding.weight.shape[0] == 51864)
+            splits = self.token_embedding.weight.split(self.token_embedding.weight.shape[0]//12, dim=0)
+            logits = torch.cat([torch.einsum('bid,jd->bij', x, split) for split in splits]).view(*x.shape[:2], -1)
+
+        return logits
+
+class WhisperANE(Whisper):
+    def __init__(self, dims: ModelDimensions):
+        super().__init__(dims)
+
+        setattr(self, 'encoder', AudioEncoderANE(
+            self.dims.n_mels,
+            self.dims.n_audio_ctx,
+            self.dims.n_audio_state,
+            self.dims.n_audio_head,
+            self.dims.n_audio_layer,
+        ))
+        setattr(self, 'decoder', TextDecoderANE(
+            self.dims.n_vocab,
+            self.dims.n_text_ctx,
+            self.dims.n_text_state,
+            self.dims.n_text_head,
+            self.dims.n_text_layer,
+        ))
+
+        self._register_load_state_dict_pre_hook(linear_to_conv2d_map)
+
+    def forward(self, mel: torch.Tensor, tokens: torch.Tensor) -> Dict[str, torch.Tensor]:
+        return self.decoder(tokens, self.encoder(mel))
+
+    def install_kv_cache_hooks(self, cache: Optional[dict] = None):
+        cache = {**cache} if cache is not None else {}
+        hooks = []
+
+        def save_to_cache(module, _, output):
+            if module not in cache or output.shape[3] > self.decoder.positional_embedding.shape[0]:
+                cache[module] = output  # save as-is, for the first token or cross attention
+            else:
+                cache[module] = torch.cat([cache[module], output], dim=3).detach()
+            return cache[module]
+
+        def install_hooks(layer: nn.Module):
+            if isinstance(layer, MultiHeadAttentionANE):
+                hooks.append(layer.key.register_forward_hook(save_to_cache))
+                hooks.append(layer.value.register_forward_hook(save_to_cache))
+
+        self.decoder.apply(install_hooks)
+        return cache, hooks
+
+def convert_encoder(hparams, model, quantize=False):
+    model.eval()
+
+    input_shape = (1, 80, 3000)
+    input_data = torch.randn(input_shape)
+    traced_model = torch.jit.trace(model, input_data)
+
+    model = ct.convert(
+        traced_model,
+        convert_to=None if quantize else "mlprogram", # convert will fail if weights are quantized, not sure why
+        inputs=[ct.TensorType(name="logmel_data", shape=input_shape)],
+        outputs=[ct.TensorType(name="output")],
+        compute_units=ct.ComputeUnit.ALL
+    )
+
+    if quantize:
+        model = quantize_weights(model, nbits=16)
+
+    return model
+
+def convert_decoder(hparams, model, quantize=False):
+    model.eval()
+
+    tokens_shape = (1, 1)
+    audio_shape = (1, hparams.n_audio_state, 1, 1500)
+
+    audio_data = torch.randn(audio_shape)
+    token_data = torch.randint(50257, tokens_shape).long()
+    traced_model = torch.jit.trace(model, (token_data, audio_data))
+
+    model = ct.convert(
+        traced_model,
+        convert_to=None if quantize else "mlprogram", # convert will fail if weights are quantized, not sure why
+        inputs=[
+            ct.TensorType(name="token_data", shape=tokens_shape, dtype=int),
+            ct.TensorType(name="audio_data", shape=audio_shape)
+        ]
+    )
+
+    if quantize:
+        model = quantize_weights(model, nbits=16)
+
+    return model
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model", type=str, help="model to convert (e.g. tiny, tiny.en, base, base.en, small, small.en, medium, medium.en, large)", required=True)
+    parser.add_argument("--encoder-only", type=bool, help="only convert encoder", default=False)
+    parser.add_argument("--quantize",     type=bool, help="quantize weights to F16", default=False)
+    parser.add_argument("--optimize-ane", type=bool, help="optimize for ANE execution (currently broken)", default=False)
+    args = parser.parse_args()
+
+    if args.model not in ["tiny", "tiny.en", "base", "base.en", "small", "small.en", "medium", "medium.en", "large"]:
+        raise ValueError("Invalid model name")
+
+    whisper = load_model(args.model).cpu()
+    hparams = whisper.dims
+    print(hparams)
+
+    if args.optimize_ane:
+        whisperANE = WhisperANE(hparams).eval()
+        whisperANE.load_state_dict(whisper.state_dict())
+
+        encoder = whisperANE.encoder
+        decoder = whisperANE.decoder
+    else:
+        encoder = whisper.encoder
+        decoder = whisper.decoder
+
+    # Convert encoder
+    encoder = convert_encoder(hparams, encoder, quantize=args.quantize)
+    encoder.save(f"models/coreml-encoder-{args.model}.mlpackage")
+
+    if args.encoder_only is False:
+        # Convert decoder
+        decoder = convert_decoder(hparams, decoder, quantize=args.quantize)
+        decoder.save(f"models/coreml-decoder-{args.model}.mlpackage")
+
+    print("done converting")

models/download-coreml-model.sh

@@ -0,0 +1,82 @@
+#!/bin/bash
+
+# This script downloads Whisper model files that have already been converted to Core ML format.
+# This way you don't have to convert them yourself.
+
+src="https://huggingface.co/datasets/ggerganov/whisper.cpp-coreml"
+pfx="resolve/main/ggml"
+
+# get the path of this script
+function get_script_path() {
+    if [ -x "$(command -v realpath)" ]; then
+        echo "$(dirname $(realpath $0))"
+    else
+        local ret="$(cd -- "$(dirname "$0")" >/dev/null 2>&1 ; pwd -P)"
+        echo "$ret"
+    fi
+}
+
+models_path="$(get_script_path)"
+
+# Whisper models
+models=( "tiny.en" "tiny" "base.en" "base" "small.en" "small" "medium.en" "medium" "large-v1" "large" )
+
+# list available models
+function list_models {
+    printf "\n"
+    printf "  Available models:"
+    for model in "${models[@]}"; do
+        printf " $model"
+    done
+    printf "\n\n"
+}
+
+if [ "$#" -ne 1 ]; then
+    printf "Usage: $0 <model>\n"
+    list_models
+
+    exit 1
+fi
+
+model=$1
+
+if [[ ! " ${models[@]} " =~ " ${model} " ]]; then
+    printf "Invalid model: $model\n"
+    list_models
+
+    exit 1
+fi
+
+# download Core ML model
+
+printf "Downloading Core ML model $model from '$src' ...\n"
+
+cd $models_path
+
+if [ -f "ggml-$model.mlmodel" ]; then
+    printf "Model $model already exists. Skipping download.\n"
+    exit 0
+fi
+
+if [ -x "$(command -v wget)" ]; then
+    wget --quiet --show-progress -O ggml-$model.mlmodel $src/$pfx-$model.mlmodel
+elif [ -x "$(command -v curl)" ]; then
+    curl -L --output ggml-$model.mlmodel $src/$pfx-$model.mlmodel
+else
+    printf "Either wget or curl is required to download models.\n"
+    exit 1
+fi
+
+
+if [ $? -ne 0 ]; then
+    printf "Failed to download Core ML model $model \n"
+    printf "Please try again later or download the original Whisper model files and convert them yourself.\n"
+    exit 1
+fi
+
+printf "Done! Model '$model' saved in 'models/ggml-$model.mlmodel'\n"
+printf "Run the following command to compile it:\n\n"
+printf "  $ xcrun coremlc compile ./models/ggml-$model.mlmodel ./models\n\n"
+printf "You can now use it like this:\n\n"
+printf "  $ ./main -m models/ggml-$model.bin -f samples/jfk.wav\n"
+printf "\n"

models/download-ggml-model.cmd

@@ -40,7 +40,7 @@ if exist "ggml-%model%.bin" (
   goto :eof
 )
 
-PowerShell -NoProfile -ExecutionPolicy Bypass -Command "Invoke-WebRequest -Uri https://huggingface.co/datasets/ggerganov/whisper.cpp/resolve/main/ggml-%model%.bin -OutFile ggml-%model%.bin"
+PowerShell -NoProfile -ExecutionPolicy Bypass -Command "Invoke-WebRequest -Uri https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-%model%.bin -OutFile ggml-%model%.bin"
 
 if %ERRORLEVEL% neq 0 (
   echo Failed to download ggml model %model%

models/download-ggml-model.sh

@@ -6,13 +6,13 @@
 #src="https://ggml.ggerganov.com"
 #pfx="ggml-model-whisper"
 
-src="https://huggingface.co/datasets/ggerganov/whisper.cpp"
+src="https://huggingface.co/ggerganov/whisper.cpp"
 pfx="resolve/main/ggml"
 
 # get the path of this script
 function get_script_path() {
     if [ -x "$(command -v realpath)" ]; then
-        echo "$(dirname $(realpath $0))"
+        echo "$(dirname "$(realpath "$0")")"
     else
         local ret="$(cd -- "$(dirname "$0")" >/dev/null 2>&1 ; pwd -P)"
         echo "$ret"

models/generate-coreml-interface.sh

@@ -0,0 +1,29 @@
+#!/bin/bash
+#
+# This generates:
+#   - coreml/whisper-encoder-impl.h and coreml/whisper-encoder-impl.m
+#   - coreml/whisper-decoder-impl.h and coreml/whisper-decoder-impl.m
+#
+
+wd=$(dirname "$0")
+cd "$wd/../"
+
+python3 models/convert-whisper-to-coreml.py --model tiny.en
+
+mv -v models/coreml-encoder-tiny.en.mlpackage models/whisper-encoder-impl.mlpackage
+xcrun coremlc generate models/whisper-encoder-impl.mlpackage coreml/
+mv coreml/whisper_encoder_impl.h coreml/whisper-encoder-impl.h
+mv coreml/whisper_encoder_impl.m coreml/whisper-encoder-impl.m
+sed -i '' 's/whisper_encoder_impl\.h/whisper-encoder-impl.h/g' coreml/whisper-encoder-impl.m
+sed -i '' 's/whisper_encoder_impl\.m/whisper-encoder-impl.m/g' coreml/whisper-encoder-impl.m
+sed -i '' 's/whisper_encoder_impl\.h/whisper-encoder-impl.h/g' coreml/whisper-encoder-impl.h
+
+mv -v models/coreml-decoder-tiny.en.mlpackage models/whisper-decoder-impl.mlpackage
+xcrun coremlc generate models/whisper-decoder-impl.mlpackage coreml/
+mv coreml/whisper_decoder_impl.h coreml/whisper-decoder-impl.h
+mv coreml/whisper_decoder_impl.m coreml/whisper-decoder-impl.m
+sed -i '' 's/whisper_decoder_impl\.h/whisper-decoder-impl.h/g' coreml/whisper-decoder-impl.m
+sed -i '' 's/whisper_decoder_impl\.m/whisper-decoder-impl.m/g' coreml/whisper-decoder-impl.m
+sed -i '' 's/whisper_decoder_impl\.h/whisper-decoder-impl.h/g' coreml/whisper-decoder-impl.h
+
+rm -rfv models/whisper-encoder-impl.mlpackage models/whisper-decoder-impl.mlpackage

models/generate-coreml-model.sh

@@ -0,0 +1,25 @@
+#!/bin/bash
+
+# Usage: ./generate-coreml-model.sh <model-name>
+if [ $# -eq 0 ]
+  then
+    echo "No model name supplied"
+    echo "Usage: ./generate-coreml-model.sh <model-name>"
+    exit 1
+fi
+
+mname="$1"
+
+wd=$(dirname "$0")
+cd "$wd/../"
+
+python3 models/convert-whisper-to-coreml.py --model $mname --encoder-only True
+
+xcrun coremlc compile models/coreml-encoder-${mname}.mlpackage models/
+rm -rf models/ggml-${mname}-encoder.mlmodelc
+mv -v models/coreml-encoder-${mname}.mlmodelc models/ggml-${mname}-encoder.mlmodelc
+
+# TODO: decoder (sometime in the future maybe)
+#xcrun coremlc compile models/whisper-decoder-${mname}.mlpackage models/
+#rm -rf models/ggml-${mname}-decoder.mlmodelc
+#mv -v models/coreml_decoder_${mname}.mlmodelc models/ggml-${mname}-decoder.mlmodelc