release : v1.5.2

* wchess: whisper assisted chess

* wchess: fix allowed moves in check

* wchess: touchstart, touchend events

* wchess: css, disabled button

* wchess : html touches

* wchess : minor fixes and code style

* wchess : bump encoder context to 1280

* wchess : index.html

* wchess : fix CI warnings

* wchess : add array header

* wchess : build static library

* wchess : display grammar

* wchess : update UX

* wchess : add comment

* wchess : add README

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

.github/workflows/build.yml

@@ -25,6 +25,7 @@ jobs:
           docker run --platform ${{ matrix.arch }} --rm \
             -v ${{ github.workspace }}:/workspace \
             -w /workspace ${{ env.ubuntu_image }} /bin/sh -c '
+            set -e
             apt update
             apt install -y build-essential libsdl2-dev
             make
@@ -86,6 +87,7 @@ jobs:
           docker run --platform ${{ matrix.arch }} --rm \
             -v ${{ github.workspace }}:/workspace \
             -w /workspace ${{ env.ubuntu_image }} /bin/sh -c '
+            set -e
             apt update
             apt install -y build-essential cmake libsdl2-dev
             cmake . -DWHISPER_SDL2=ON -DCMAKE_BUILD_TYPE=${{ matrix.build }}
@@ -113,8 +115,10 @@ jobs:
           docker run --platform ${{ matrix.arch }} --rm \
             -v ${{ github.workspace }}:/workspace \
             -w /workspace ${{ env.ubuntu_image }} /bin/sh -c '
+            set -e
             apt update
-            apt install -y build-essential cmake libsdl2-dev
+            apt install -y clang
+            apt install -y clang build-essential cmake libsdl2-dev
             cmake . -DWHISPER_SDL2=ON -DCMAKE_BUILD_TYPE=${{ matrix.build }} -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_C_COMPILER=clang
             make
             ctest -L gh --output-on-failure'
@@ -140,6 +144,7 @@ jobs:
           docker run --platform ${{ matrix.arch }} --rm \
             -v ${{ github.workspace }}:/workspace \
             -w /workspace ${{ env.ubuntu_image }} /bin/sh -c '
+            set -e
             apt update
             apt install -y build-essential cmake
             cmake . -DCMAKE_BUILD_TYPE=Debug -DWHISPER_SANITIZE_${{ matrix.sanitizer }}=ON
@@ -217,10 +222,10 @@ jobs:
         sdl2: [ON]
         include:
           - arch: Win32
-            obzip: https://github.com/OpenMathLib/OpenBLAS/releases/download/v0.3.24/OpenBLAS-0.3.24-x86.zip
+            obzip: https://github.com/OpenMathLib/OpenBLAS/releases/download/v0.3.25/OpenBLAS-0.3.25-x86.zip
             s2arc: x86
           - arch: x64
-            obzip: https://github.com/OpenMathLib/OpenBLAS/releases/download/v0.3.24/OpenBLAS-0.3.24-x64.zip
+            obzip: https://github.com/OpenMathLib/OpenBLAS/releases/download/v0.3.25/OpenBLAS-0.3.25-x64.zip
             s2arc: x64
           - sdl2: ON
             s2ver: 2.26.0
@@ -285,6 +290,7 @@ jobs:
         arch: [x64]
         cublas: [ON]
         sdl2: [ON]
+        cuda-toolkit: [12.2.0, 11.8.0]
         include:
           - arch: x64
             s2arc: x64
@@ -300,7 +306,9 @@ jobs:
 
       - name: Install CUDA Toolkit
         id: cuda-toolkit
-        uses: Jimver/cuda-toolkit@v0.2.10
+        uses: Jimver/cuda-toolkit@v0.2.11
+        with:
+          cuda: '${{ matrix.cuda-toolkit }}'
 
       - name: Fetch SDL2 and set SDL2_DIR
         if: matrix.sdl2 == 'ON'
@@ -315,10 +323,10 @@ jobs:
           -DCMAKE_BUILD_TYPE=${{ matrix.build }}
           -DWHISPER_CUBLAS=1
 
-      - name: Build
+      - name: Build ${{ matrix.cuda-toolkit }}
         run: |
           cd ./build
-          msbuild ALL_BUILD.vcxproj -t:build -p:configuration=${{ matrix.build }} -p:platform=${{ matrix.arch }}
+          cmake --build . --config ${{ matrix.build }}
 
       - name: Copy CUDA DLLs
         run: >
@@ -335,7 +343,7 @@ jobs:
         if: matrix.sdl2 == 'ON'
         uses: actions/upload-artifact@v1
         with:
-          name: whisper-cublas-bin-${{ matrix.arch }}
+          name: whisper-cublas-${{ matrix.cuda-toolkit }}-bin-${{ matrix.arch }}
           path: build/bin/${{ matrix.build }}
 
   emscripten:

CMakeLists.txt

@@ -1,6 +1,6 @@
 cmake_minimum_required (VERSION 3.5)
 
-project(whisper.cpp VERSION 1.5.0)
+project(whisper.cpp VERSION 1.5.2)
 
 # Add path to modules
 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
@@ -533,7 +533,7 @@ target_compile_definitions(${TARGET} PUBLIC
     ${WHISPER_EXTRA_FLAGS}
     )
 
-set_target_properties(${TARGET} PROPERTIES PUBLIC_HEADER "whisper.h")
+set_target_properties(${TARGET} PROPERTIES PUBLIC_HEADER "ggml.h;whisper.h")
 
 include(GNUInstallDirs)
 

Package.swift

@@ -2,33 +2,14 @@
 
 import PackageDescription
 
-#if arch(arm) || arch(arm64)
+let package = Package(
-let platforms: [SupportedPlatform]? = [
+    name: "whisper",
+    platforms: [
         .macOS(.v12),
         .iOS(.v14),
         .watchOS(.v4),
         .tvOS(.v14)
-]
+    ],
-let exclude: [String] = []
-let resources: [Resource] = [
-    .process("ggml-metal.metal")
-]
-let additionalSources: [String] = ["ggml-metal.m"]
-let additionalSettings: [CSetting] = [
-    .unsafeFlags(["-fno-objc-arc"]),
-    .define("GGML_USE_METAL")
-]
-#else
-let platforms: [SupportedPlatform]? = nil
-let exclude: [String] = ["ggml-metal.metal"]
-let resources: [Resource] = []
-let additionalSources: [String] = []
-let additionalSettings: [CSetting] = []
-#endif
-
-let package = Package(
-    name: "whisper",
-    platforms: platforms,
     products: [
         .library(name: "whisper", targets: ["whisper"]),
     ],
@@ -36,7 +17,7 @@ let package = Package(
         .target(
             name: "whisper",
             path: ".",
-            exclude: exclude + [
+            exclude: [
                "bindings",
                "cmake",
                "coreml",
@@ -55,19 +36,22 @@ let package = Package(
                 "whisper.cpp",
                 "ggml-alloc.c",
                 "ggml-backend.c",
-                "ggml-quants.c"
+                "ggml-quants.c",
-            ] + additionalSources,
+                "ggml-metal.m"
-            resources: resources,
+            ],
+            resources: [.process("ggml-metal.metal")],
             publicHeadersPath: "spm-headers",
             cSettings: [
                 .unsafeFlags(["-Wno-shorten-64-to-32", "-O3", "-DNDEBUG"]),
-                .define("GGML_USE_ACCELERATE")
+                .define("GGML_USE_ACCELERATE"),
+                .unsafeFlags(["-fno-objc-arc"]),
+                .define("GGML_USE_METAL")
                 // NOTE: NEW_LAPACK will required iOS version 16.4+
                 // We should consider add this in the future when we drop support for iOS 14
                 // (ref: ref: https://developer.apple.com/documentation/accelerate/1513264-cblas_sgemm?language=objc)
                 // .define("ACCELERATE_NEW_LAPACK"),
                 // .define("ACCELERATE_LAPACK_ILP64")
-            ] + additionalSettings,
+            ],
             linkerSettings: [
                 .linkedFramework("Accelerate")
             ]

README.md

@@ -6,7 +6,7 @@
 [![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: [v1.5.0](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.5.0) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)
+Stable: [v1.5.2](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.5.2) / [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:
 
@@ -110,8 +110,8 @@ options:
   -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
   -sow,      --split-on-word     [false  ] split on word rather than on token
-  -bo N,     --best-of N         [2      ] number of best candidates to keep
+  -bo N,     --best-of N         [5      ] number of best candidates to keep
-  -bs N,     --beam-size N       [-1     ] beam size for beam search
+  -bs N,     --beam-size N       [5      ] 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
@@ -128,6 +128,7 @@ options:
   -fp,       --font-path         [/System/Library/Fonts/Supplemental/Courier New Bold.ttf] path to a monospace font for karaoke video
   -ocsv,     --output-csv        [false  ] output result in a CSV file
   -oj,       --output-json       [false  ] output result in a JSON file
+  -ojf,      --output-json-full  [false  ] include more information in the 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
@@ -139,7 +140,8 @@ options:
   -m FNAME,  --model FNAME       [models/ggml-base.en.bin] model path
   -f FNAME,  --file FNAME        [       ] input WAV file path
   -oved D,   --ov-e-device DNAME [CPU    ] the OpenVINO device used for encode inference
-  -ls,       --log-score         [false  ] log best decoder scores of token
+  -ls,       --log-score         [false  ] log best decoder scores of tokens
+  -ng,       --no-gpu            [false  ] disable GPU
 
 
 bash ./models/download-ggml-model.sh base.en
@@ -768,6 +770,7 @@ Some of the examples are even ported to run in the browser using WebAssembly. Ch
 | [bench](examples/bench) | [bench.wasm](examples/bench.wasm) | Benchmark the performance of Whisper on your machine |
 | [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 |
+| [wchess](examples/wchess) | [wchess.wasm](examples/wchess) | Voice-controlled chess |
 | [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 |
@@ -777,6 +780,7 @@ Some of the examples are even ported to run in the browser using WebAssembly. Ch
 | [generate-karaoke.sh](examples/generate-karaoke.sh) | | Helper script to easily [generate a karaoke video](https://youtu.be/uj7hVta4blM) of raw audio capture |
 | [livestream.sh](examples/livestream.sh) | | [Livestream audio transcription](https://github.com/ggerganov/whisper.cpp/issues/185) |
 | [yt-wsp.sh](examples/yt-wsp.sh) | | Download + transcribe and/or translate any VOD [(original)](https://gist.github.com/DaniruKun/96f763ec1a037cc92fe1a059b643b818) |
+| [server](examples/server) | | HTTP transcription server with OAI-like API |
 
 ## [Discussions](https://github.com/ggerganov/whisper.cpp/discussions)
 

bindings/go/Makefile

@@ -1,9 +1,26 @@
+ifndef UNAME_S
+UNAME_S := $(shell uname -s)
+endif
+
+ifndef UNAME_P
+UNAME_P := $(shell uname -p)
+endif
+
+ifndef UNAME_M
+UNAME_M := $(shell uname -m)
+endif
+
+GGML_METAL_PATH_RESOURCES := $(abspath ../..)
 BUILD_DIR := build
 MODELS_DIR := models
 EXAMPLES_DIR := $(wildcard examples/*)
 INCLUDE_PATH := $(abspath ../..)
 LIBRARY_PATH := $(abspath ../..)
 
+ifeq ($(UNAME_S),Darwin)
+	EXT_LDFLAGS := -framework Foundation -framework Metal -framework MetalKit
+endif
+
 all: clean whisper examples
 
 whisper: mkdir
@@ -11,8 +28,13 @@ whisper: mkdir
 	@${MAKE} -C ../.. libwhisper.a
 
 test: model-small whisper modtidy
+ifeq ($(UNAME_S),Darwin)
+	@C_INCLUDE_PATH=${INCLUDE_PATH} LIBRARY_PATH=${LIBRARY_PATH} GGML_METAL_PATH_RESOURCES=${GGML_METAL_PATH_RESOURCES} go test -ldflags "-extldflags '$(EXT_LDFLAGS)'" -v .
+	@C_INCLUDE_PATH=${INCLUDE_PATH} LIBRARY_PATH=${LIBRARY_PATH} GGML_METAL_PATH_RESOURCES=${GGML_METAL_PATH_RESOURCES} go test -ldflags "-extldflags '$(EXT_LDFLAGS)'" -v ./pkg/whisper/...
+else
 	@C_INCLUDE_PATH=${INCLUDE_PATH} LIBRARY_PATH=${LIBRARY_PATH} go test -v .
 	@C_INCLUDE_PATH=${INCLUDE_PATH} LIBRARY_PATH=${LIBRARY_PATH} go test -v ./pkg/whisper/...
+endif
 
 examples: $(EXAMPLES_DIR)
 
@@ -21,7 +43,11 @@ model-small: mkdir examples/go-model-download
 
 $(EXAMPLES_DIR): mkdir whisper modtidy
 	@echo Build example $(notdir $@)
+ifeq ($(UNAME_S),Darwin)
+	@C_INCLUDE_PATH=${INCLUDE_PATH} LIBRARY_PATH=${LIBRARY_PATH} GGML_METAL_PATH_RESOURCES=${GGML_METAL_PATH_RESOURCES} go build ${BUILD_FLAGS} -ldflags "-extldflags '$(EXT_LDFLAGS)'" -o ${BUILD_DIR}/$(notdir $@) ./$@
+else
 	@C_INCLUDE_PATH=${INCLUDE_PATH} LIBRARY_PATH=${LIBRARY_PATH} go build ${BUILD_FLAGS} -o ${BUILD_DIR}/$(notdir $@) ./$@
+endif
 
 mkdir:
 	@echo Mkdir ${BUILD_DIR}

bindings/javascript/package.json

@@ -1,6 +1,6 @@
 {
   "name": "whisper.cpp",
-  "version": "1.5.0",
+  "version": "1.5.2",
   "description": "Whisper speech recognition",
   "main": "whisper.js",
   "scripts": {

examples/CMakeLists.txt

@@ -73,3 +73,5 @@ else()
     add_subdirectory(talk-llama)
     add_subdirectory(lsp)
 endif()
+
+add_subdirectory(wchess)

examples/helpers.js

@@ -22,6 +22,7 @@ var printTextarea = (function() {
 async function clearCache() {
     if (confirm('Are you sure you want to clear the cache?\nAll the models will be downloaded again.')) {
         indexedDB.deleteDatabase(dbName);
+        location.reload();
     }
 }
 

examples/main/README.md

@@ -17,28 +17,37 @@ options:
   -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
+  -sow,      --split-on-word     [false  ] split on word rather than on token
   -bo N,     --best-of N         [5      ] number of best candidates to keep
-  -bs N,     --beam-size N       [-1     ] beam size for beam search
+  -bs N,     --beam-size N       [5      ] 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)
+  -debug,    --debug-mode        [false  ] enable debug mode (eg. dump log_mel)
   -tr,       --translate         [false  ] translate from source language to english
   -di,       --diarize           [false  ] stereo audio diarization
+  -tdrz,     --tinydiarize       [false  ] enable tinydiarize (requires a tdrz model)
   -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
+  -olrc,     --output-lrc        [false  ] output result in a lrc file
   -owts,     --output-words      [false  ] output script for generating karaoke video
+  -fp,       --font-path         [/System/Library/Fonts/Supplemental/Courier New Bold.ttf] path to a monospace font for karaoke video
   -ocsv,     --output-csv        [false  ] output result in a CSV file
   -oj,       --output-json       [false  ] output result in a JSON file
+  -ojf,      --output-json-full  [false  ] include more information in the 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
+  -nt,       --no-timestamps     [false  ] do not print timestamps
   -l LANG,   --language LANG     [en     ] spoken language ('auto' for auto-detect)
+  -dl,       --detect-language   [false  ] exit after automatically detecting language
              --prompt PROMPT     [       ] initial prompt
   -m FNAME,  --model FNAME       [models/ggml-base.en.bin] model path
   -f FNAME,  --file FNAME        [       ] input WAV file path
+  -oved D,   --ov-e-device DNAME [CPU    ] the OpenVINO device used for encode inference
+  -ls,       --log-score         [false  ] log best decoder scores of tokens
+  -ng,       --no-gpu            [false  ] disable GPU
 ```

examples/server/CMakeLists.txt

@@ -4,3 +4,9 @@ add_executable(${TARGET} server.cpp httplib.h json.hpp)
 include(DefaultTargetOptions)
 
 target_link_libraries(${TARGET} PRIVATE common whisper ${CMAKE_THREAD_LIBS_INIT})
+
+# Check if the compiler is MinGW
+if(MINGW)
+    # Link the necessary libraries for SSL and Winsock
+    target_link_libraries(${TARGET} PRIVATE -lcrypt32 -lssl -lcrypto -lws2_32)
+endif()

examples/server/README.md

@@ -2,6 +2,10 @@
 
 Simple http server. WAV Files are passed to the inference model via http requests.
 
+https://github.com/ggerganov/whisper.cpp/assets/1991296/e983ee53-8741-4eb5-9048-afe5e4594b8f
+
+## Usage
+
 ```
 ./server -h
 
@@ -29,6 +33,7 @@ options:
   -nf,       --no-fallback       [false  ] do not use temperature fallback while decoding
   -ps,       --print-special     [false  ] print special tokens
   -pc,       --print-colors      [false  ] print colors
+  -pr,       --print-realtime    [false  ] print output in realtime
   -pp,       --print-progress    [false  ] print progress
   -nt,       --no-timestamps     [false  ] do not print timestamps
   -l LANG,   --language LANG     [en     ] spoken language ('auto' for auto-detect)
@@ -38,8 +43,12 @@ options:
   -oved D,   --ov-e-device DNAME [CPU    ] the OpenVINO device used for encode inference
   --host HOST,                   [127.0.0.1] Hostname/ip-adress for the server
   --port PORT,                   [8080   ] Port number for the server
+  --convert,                     [false  ] Convert audio to WAV, requires ffmpeg on the server
 ```
 
+> [!WARNING]  
+> **Do not run the server example with administrative privileges and ensure it's operated in a sandbox environment, especially since it involves risky operations like accepting user file uploads and using ffmpeg for format conversions. Always validate and sanitize inputs to guard against potential security threats.**
+
 ## request examples
 
 **/inference**

examples/server/server.cpp

@@ -11,6 +11,7 @@
 #include <thread>
 #include <vector>
 #include <cstring>
+#include <sstream>
 
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
@@ -43,6 +44,8 @@ struct server_params
     int32_t port          = 8080;
     int32_t read_timeout  = 600;
     int32_t write_timeout = 600;
+    
+    bool ffmpeg_converter = false;
 };
 
 struct whisper_params {
@@ -72,6 +75,7 @@ struct whisper_params {
     bool no_fallback     = false;
     bool print_special   = false;
     bool print_colors    = false;
+    bool print_realtime  = false;
     bool print_progress  = false;
     bool no_timestamps   = false;
     bool use_gpu         = true;
@@ -144,6 +148,7 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     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, "  -pc,       --print-colors      [%-7s] print colors\n",                                   params.print_colors ? "true" : "false");
+    fprintf(stderr, "  -pr,       --print-realtime    [%-7s] print output in realtime\n",                       params.print_realtime ? "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 ? "true" : "false");
     fprintf(stderr, "  -l LANG,   --language LANG     [%-7s] spoken language ('auto' for auto-detect)\n",       params.language.c_str());
@@ -155,6 +160,7 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  --host HOST,                   [%-7s] Hostname/ip-adress for the server\n", sparams.hostname.c_str());
     fprintf(stderr, "  --port PORT,                   [%-7d] Port number for the server\n", sparams.port);
     fprintf(stderr, "  --public PATH,                 [%-7s] Path to the public folder\n", sparams.public_path.c_str());
+    fprintf(stderr, "  --convert,                     [%-7s] Convert audio to WAV, requires ffmpeg on the server", sparams.ffmpeg_converter ? "true" : "false");
     fprintf(stderr, "\n");
 }
 
@@ -188,6 +194,7 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params, serve
         else if (arg == "-fp"   || arg == "--font-path")       { params.font_path       = 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 == "-pr"   || arg == "--print-realtime")  { params.print_realtime  = true; }
         else if (arg == "-pp"   || arg == "--print-progress")  { params.print_progress  = true; }
         else if (arg == "-nt"   || arg == "--no-timestamps")   { params.no_timestamps   = true; }
         else if (arg == "-l"    || arg == "--language")        { params.language        = argv[++i]; }
@@ -200,6 +207,7 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params, serve
         else if (                  arg == "--port")            { sparams.port        = std::stoi(argv[++i]); }
         else if (                  arg == "--host")            { sparams.hostname    = argv[++i]; }
         else if (                  arg == "--public")          { sparams.public_path = argv[++i]; }
+        else if (                  arg == "--convert")         { sparams.ffmpeg_converter     = true; }
         else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             whisper_print_usage(argc, argv, params, sparams);
@@ -217,6 +225,45 @@ struct whisper_print_user_data {
     int progress_prev;
 };
 
+void check_ffmpeg_availibility() {
+    int result = system("ffmpeg -version");
+
+    if (result == 0) {
+        std::cout << "ffmpeg is available." << std::endl;
+    } else {
+        // ffmpeg is not available
+        std::cout << "ffmpeg is not found. Please ensure that ffmpeg is installed ";
+        std::cout << "and that its executable is included in your system's PATH. ";
+        exit(0);
+    }
+}
+
+bool convert_to_wav(const std::string & temp_filename, std::string & error_resp) {
+    std::ostringstream cmd_stream;
+    std::string converted_filename_temp = temp_filename + "_temp.wav";
+    cmd_stream << "ffmpeg -i \"" << temp_filename << "\" -ar 16000 -ac 1 -c:a pcm_s16le \"" << converted_filename_temp << "\" 2>&1";
+    std::string cmd = cmd_stream.str();
+
+    int status = std::system(cmd.c_str());
+    if (status != 0) {
+        error_resp = "{\"error\":\"FFmpeg conversion failed.\"}";
+        return false;
+    }
+
+    // Remove the original file
+    if (remove(temp_filename.c_str()) != 0) {
+        error_resp = "{\"error\":\"Failed to remove the original file.\"}";
+        return false;
+    }
+
+    // Rename the temporary file to match the original filename
+    if (rename(converted_filename_temp.c_str(), temp_filename.c_str()) != 0) {
+        error_resp = "{\"error\":\"Failed to rename the temporary file.\"}";
+        return false;
+    }
+    return true;
+}
+
 std::string estimate_diarization_speaker(std::vector<std::vector<float>> pcmf32s, int64_t t0, int64_t t1, bool id_only = false) {
     std::string speaker = "";
     const int64_t n_samples = pcmf32s[0].size();
@@ -373,7 +420,7 @@ void get_req_parameters(const Request & req, whisper_params & params)
     {
         params.response_format = req.get_file_value("response-format").content;
     }
-    if (req.has_file("temerature"))
+    if (req.has_file("temperature"))
     {
         params.userdef_temp = std::stof(req.get_file_value("temperature").content);
     }
@@ -404,6 +451,9 @@ int main(int argc, char ** argv) {
         exit(0);
     }
 
+    if (sparams.ffmpeg_converter) {
+        check_ffmpeg_availibility();
+    }
     // whisper init
     struct whisper_context_params cparams;
     cparams.use_gpu = params.use_gpu;
@@ -419,6 +469,9 @@ int main(int argc, char ** argv) {
     whisper_ctx_init_openvino_encoder(ctx, nullptr, params.openvino_encode_device.c_str(), nullptr);
 
     Server svr;
+    svr.set_default_headers({{"Server", "whisper.cpp"},
+                             {"Access-Control-Allow-Origin", "*"},
+                             {"Access-Control-Allow-Headers", "content-type"}});
 
     std::string const default_content = "<html>hello</html>";
 
@@ -429,7 +482,7 @@ int main(int argc, char ** argv) {
     });
 
     svr.Post("/inference", [&](const Request &req, Response &res){
-        // aquire whisper model mutex lock
+        // acquire whisper model mutex lock
         whisper_mutex.lock();
 
         // first check user requested fields of the request
@@ -453,20 +506,35 @@ int main(int argc, char ** argv) {
         std::vector<float> pcmf32;               // mono-channel F32 PCM
         std::vector<std::vector<float>> pcmf32s; // stereo-channel F32 PCM
 
-        // write file to temporary file
+        // write to temporary file
-        std::ofstream temp_file{filename, std::ios::binary};
+        const std::string temp_filename = "whisper_server_temp_file.wav";
+        std::ofstream temp_file{temp_filename, std::ios::binary};
         temp_file << audio_file.content;
+        temp_file.close();
 
-        // read wav content into pcmf32
+        // if file is not wav, convert to wav
-        if (!::read_wav(filename, pcmf32, pcmf32s, params.diarize)) {
+        
-            fprintf(stderr, "error: failed to read WAV file '%s'\n", filename.c_str());
+        if (sparams.ffmpeg_converter) {
-            const std::string error_resp = "{\"error\":\"failed to read WAV file\"}";
+            std::string error_resp = "{\"error\":\"Failed to execute ffmpeg command.\"}";
+            const bool is_converted = convert_to_wav(temp_filename, error_resp);
+            if (!is_converted) {
                 res.set_content(error_resp, "application/json");
                 whisper_mutex.unlock();
                 return;
             }
+        }
+
+        // read wav content into pcmf32
+        if (!::read_wav(temp_filename, pcmf32, pcmf32s, params.diarize)) {
+            fprintf(stderr, "error: failed to read WAV file '%s'\n", temp_filename.c_str());
+            const std::string error_resp = "{\"error\":\"failed to read WAV file\"}";
+            res.set_content(error_resp, "application/json");
+            std::remove(temp_filename.c_str());
+            whisper_mutex.unlock();
+            return;
+        }
         // remove temp file
-        std::remove(filename.c_str());
+        std::remove(temp_filename.c_str());
 
         printf("Successfully loaded %s\n", filename.c_str());
 
@@ -503,7 +571,6 @@ int main(int argc, char ** argv) {
 
         // run the inference
         {
-
             printf("Running whisper.cpp inference on %s\n", filename.c_str());
             whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
 
@@ -522,6 +589,7 @@ int main(int argc, char ** argv) {
             wparams.duration_ms      = params.duration_ms;
 
             wparams.thold_pt         = params.word_thold;
+            wparams.max_len          = params.max_len == 0 ? 60 : params.max_len;
             wparams.split_on_word    = params.split_on_word;
 
             wparams.speed_up         = params.speed_up;
@@ -541,7 +609,7 @@ int main(int argc, char ** argv) {
             whisper_print_user_data user_data = { &params, &pcmf32s, 0 };
 
             // this callback is called on each new segment
-            if (!wparams.print_realtime) {
+            if (params.print_realtime) {
                 wparams.new_segment_callback           = whisper_print_segment_callback;
                 wparams.new_segment_callback_user_data = &user_data;
             }
@@ -591,6 +659,50 @@ int main(int argc, char ** argv) {
             std::string results = output_str(ctx, params, pcmf32s);
             res.set_content(results.c_str(), "text/html");
         }
+        else if (params.response_format == srt_format)
+        {
+            std::stringstream ss;
+            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);
+                std::string speaker = "";
+
+                if (params.diarize && pcmf32s.size() == 2)
+                {
+                    speaker = estimate_diarization_speaker(pcmf32s, t0, t1);
+                }
+
+                ss << i + 1 + params.offset_n << "\n";
+                ss << to_timestamp(t0, true) << " --> " << to_timestamp(t1, true) << "\n";
+                ss << speaker << text << "\n\n";
+            }
+            res.set_content(ss.str(), "application/x-subrip");
+        } else if (params.response_format == vtt_format) {
+            std::stringstream ss;
+
+            ss << "WEBVTT\n\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 t0 = whisper_full_get_segment_t0(ctx, i);
+                const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
+                std::string speaker = "";
+
+                if (params.diarize && pcmf32s.size() == 2)
+                {
+                    speaker = estimate_diarization_speaker(pcmf32s, t0, t1, true);
+                    speaker.insert(0, "<v Speaker");
+                    speaker.append(">");
+                }
+
+                ss << to_timestamp(t0) << " --> " << to_timestamp(t1) << "\n";
+                ss << speaker << text << "\n\n";
+            }
+            res.set_content(ss.str(), "text/vtt");
+        }
         // TODO add more output formats
         else
         {

examples/talk-llama/CMakeLists.txt

@@ -18,6 +18,11 @@ if (WHISPER_SDL2)
         ../../ggml-quants.c
         ../../whisper.cpp)
 
+    if(WIN32)
+    # It requires Windows 8.1 or later for PrefetchVirtualMemory
+    target_compile_definitions(${TARGET} PRIVATE -D_WIN32_WINNT=0x0602)
+    endif()
+
     target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS} ../../)
     target_link_libraries(${TARGET} PRIVATE ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
 

examples/wchess/CMakeLists.txt

@@ -0,0 +1,9 @@
+set(CMAKE_CXX_STANDARD 11)
+
+add_subdirectory(libwchess)
+
+if (EMSCRIPTEN)
+    add_subdirectory(wchess.wasm)
+else()
+    add_subdirectory(wchess.cmd)
+endif()

examples/wchess/README.md

@@ -0,0 +1,40 @@
+# wchess
+
+Voice-controlled chess using Whisper
+
+Online demo: https://whisper.ggerganov.com/wchess/
+
+https://github.com/ggerganov/whisper.cpp/assets/1991296/c2b2f03c-9684-49f3-8106-357d2d4e67fa
+
+## Command-line tool
+
+```bash
+mkdir build && cd build
+cmake -DWHISPER_SDL2=1 ..
+make -j
+
+./bin/wchess -m ../models/ggml-base.en.bin
+
+Move: start
+
+a b c d e f g h
+r n b q k b n r 8
+p p p p p p p p 7
+. * . * . * . * 6
+* . * . * . * . 5
+. * . * . * . * 4
+* . * . * . * . 3
+P P P P P P P P 2
+R N B Q K B N R 1
+
+White's turn
+[(l)isten/(p)ause/(q)uit]: 
+```
+
+## TODO
+
+- Improve web-browser audio capture - sometimes it does not record the voice properly
+- Add support for more languages by making the generated grammar string multi-lingual
+- Fix bugs in the chess moves logic
+
+PRs welcome!

examples/wchess/libwchess/CMakeLists.txt

@@ -0,0 +1,19 @@
+add_library(wchess-core STATIC
+    WChess.cpp
+    WChess.h
+    Chessboard.cpp
+    Chessboard.h
+)
+
+target_link_libraries(wchess-core
+    PUBLIC
+    whisper
+    common
+)
+
+target_include_directories(wchess-core
+    PUBLIC
+    "$<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}>"
+)
+
+# add_executable(test-chessboard test-chessboard.cpp Chessboard.cpp)

examples/wchess/libwchess/Chessboard.cpp

@@ -0,0 +1,803 @@
+#include "Chessboard.h"
+
+#include <array>
+#include <vector>
+#include <algorithm>
+#include <cstring>
+#include <set>
+#include <list>
+#include <chrono>
+
+namespace {
+constexpr std::array<const char*, 64> positions = {
+    "a1", "b1", "c1", "d1", "e1", "f1", "g1", "h1",
+    "a2", "b2", "c2", "d2", "e2", "f2", "g2", "h2",
+    "a3", "b3", "c3", "d3", "e3", "f3", "g3", "h3",
+    "a4", "b4", "c4", "d4", "e4", "f4", "g4", "h4",
+    "a5", "b5", "c5", "d5", "e5", "f5", "g5", "h5",
+    "a6", "b6", "c6", "d6", "e6", "f6", "g6", "h6",
+    "a7", "b7", "c7", "d7", "e7", "f7", "g7", "h7",
+    "a8", "b8", "c8", "d8", "e8", "f8", "g8", "h8",
+};
+constexpr char INVALID_POS = positions.size();
+constexpr int R = 0; // rank index
+constexpr int F = 1; // file index
+#define FILE (c[F] - '1')
+#define RANK (c[R] - 'a')
+constexpr char operator ""_P(const char * c, size_t size) {
+    return size < 2 || RANK < 0 || RANK > 7 ||
+        FILE < 0 || FILE > 7 ? INVALID_POS : FILE * 8 + RANK;
+}
+#undef FILE
+#undef RANK
+
+struct sview {
+    const char * ptr = nullptr;
+    size_t size = 0;
+
+    sview() = default;
+    sview(const char * p, size_t s) : ptr(p), size(s) {}
+    sview(const std::string& s) : ptr(s.data()), size(s.size()) {}
+
+    size_t find(char del, size_t pos) {
+        while (pos < size && ptr[pos] != del) ++pos;
+        return pos < size ? pos : std::string::npos;
+    }
+};
+
+std::vector<sview> split(sview str, char del) {
+    std::vector<sview> res;
+    size_t cur = 0;
+    size_t last = 0;
+    while (cur != std::string::npos) {
+        if (str.ptr[last] == ' ') {
+            ++last;
+            continue;
+        }
+        cur = str.find(del, last);
+        size_t len = cur == std::string::npos ? str.size - last : cur - last;
+        res.emplace_back(str.ptr + last, len);
+        last = cur + 1;
+    }
+    return res;
+}
+
+char strToPos(sview str) {
+    return operator ""_P(str.ptr, str.size);
+}
+
+constexpr std::array<const char*, 6> pieceNames =  {
+    "pawn", "knight", "bishop", "rook", "queen", "king",
+};
+
+static constexpr std::array<char, 6> blackShort =  {
+    'p', 'n', 'b', 'r', 'q', 'k',
+};
+static constexpr std::array<char, 6> whiteShort =  {
+    'P', 'N', 'B', 'R', 'Q', 'K',
+};
+
+char strToType(sview str) {
+    auto it = std::find_if(pieceNames.begin(), pieceNames.end(), [str] (const char* name) { return strncmp(name, str.ptr, str.size) == 0; });
+    return it != pieceNames.end() ? it - pieceNames.begin() : pieceNames.size();
+}
+
+// directions
+using Direction = std::array<char, 2>;
+
+constexpr Direction N   = {(char)  0, (char)  1};
+constexpr Direction NNE = {(char)  1, (char)  2};
+constexpr Direction NE  = {(char)  1, (char)  1};
+constexpr Direction ENE = {(char)  2, (char)  1};
+constexpr Direction E   = {(char)  1, (char)  0};
+constexpr Direction ESE = {(char)  2, (char) -1};
+constexpr Direction SE  = {(char)  1, (char) -1};
+constexpr Direction SSE = {(char)  1, (char) -2};
+constexpr Direction S   = {(char)  0, (char) -1};
+constexpr Direction SSW = {(char) -1, (char) -2};
+constexpr Direction SW  = {(char) -1, (char) -1};
+constexpr Direction WSW = {(char) -2, (char) -1};
+constexpr Direction W   = {(char) -1, (char)  0};
+constexpr Direction WNW = {(char) -2, (char)  1};
+constexpr Direction NW  = {(char) -1, (char)  1};
+constexpr Direction NNW = {(char) -1, (char)  2};
+
+char makeStep(char pos, const Direction& d) {
+    char next[2] = { char(positions[pos][R] + d[R]) , char(positions[pos][F] + d[F]) };
+    return strToPos(sview{next, sizeof(next)});
+}
+
+template<class Modifier>
+char traverse(char pos, const Direction& d, const Modifier& m, int count = 8) {
+    while (--count >= 0) {
+        pos = makeStep(pos, d);
+        if (pos == INVALID_POS || m(pos)) break;
+    }
+    return pos;
+}
+
+Direction normalize(const Direction& distance) {
+    //return {char((distance[R] > 0) - (distance[R] < 0)), char((distance[F] > 0) - (distance[F] < 0))};
+    const int drp = distance[R] > 0 ? 1 : 0;
+    const int drn = distance[R] < 0 ? 1 : 0;
+    const int dfp = distance[F] > 0 ? 1 : 0;
+    const int dfn = distance[F] < 0 ? 1 : 0;
+    return {char(drp - drn), char(dfp - dfn)};
+}
+
+struct Pin {
+    Direction d;
+    Piece* pinner;
+    Piece* pinned;
+};
+using Pins = std::list<Pin>;
+using Board = std::array<Piece*, 64>;
+
+std::vector<Direction> filter(const Direction& pin, std::initializer_list<Direction> directions) {
+    if (pin[R] == 0 && pin[F] == 0) return directions;
+    std::vector<Direction> result;
+    for (auto& d : directions) {
+        if ((d[R] == pin[R] || d[R] == -pin[R]) && (d[F] == pin[F] || d[F] == -pin[F])) result.push_back(d);
+    }
+    return result;
+}
+}
+
+class Piece {
+public:
+    enum Types : char {
+        Pawn,
+        Knight,
+        Bishop,
+        Rook,
+        Queen,
+        King,
+        //
+        NUM_PIECES
+    };
+
+    enum Colors : char {
+        White,
+        Black,
+    };
+
+    const char* name() const;
+    char initial() const;
+    Types type() const { return m_type; }
+    Colors color() const { return m_color; }
+    char pos() const { return m_pos; }
+    void setPos(char pos) {
+        m_pos = pos;
+        invalidate();
+    }
+    const char* coord() const;
+    const std::set<char>& allowed() const { return m_allowed; }
+    bool canReach(char pos) const;
+    virtual bool movePattern(char pos) const = 0;
+    void take();
+    virtual void reinit(const State& state) = 0;
+    void invalidate();
+protected:
+    Piece(Types type, Colors color, char pos, std::set<char> allowed)
+        : m_type(type), m_color(color), m_pos(pos), m_allowed(std::move(allowed)) {}
+    Piece(const Piece&) = delete;
+    ~Piece() = default;
+
+    const Types m_type;
+    const Colors m_color;
+    char m_pos;
+    std::set<char> m_allowed;
+    bool m_update = false;
+};
+
+struct Pawn : public Piece {
+    Pawn(Colors color, char pos, std::set<char> next) : Piece(Types::Pawn, color, pos, std::move(next)) {}
+
+    bool is_first_move() const {
+        return m_color ? coord()[F] == '7' : coord()[F] == '2';
+    }
+
+    virtual bool movePattern(char pos) const override {
+        if (m_pos == INVALID_POS) return false;
+        auto cur = coord();
+        auto next = positions[pos];
+        Direction distance = {char(next[R] - cur[R]), char(next[F] - cur[F])};
+        char forward = m_color ? -1 : 1;
+        return (forward == distance[F] && distance[R] * distance[R] <= 1)
+            || (is_first_move() && 2 * forward == distance[F] && distance[R] == 0);
+    }
+
+    virtual void reinit(const State& state) override;
+};
+
+struct Knight : public Piece {
+    Knight(Colors color, char pos, std::set<char> next) : Piece(Types::Knight, color, pos, std::move(next)) {}
+
+    virtual bool movePattern(char pos) const override {
+        if (m_pos == INVALID_POS) return false;
+        auto cur = coord();
+        auto next = positions[pos];
+        Direction diff = {char(next[R] - cur[R]), char(next[F] - cur[F])};
+        return diff[R]*diff[R] + diff[F]*diff[F] == 5;
+    }
+
+    virtual void reinit(const State& state) override;
+};
+
+struct Bishop : public Piece {
+    Bishop(Colors color, char pos) : Piece(Types::Bishop, color, pos, {}) {}
+
+    virtual bool movePattern(char pos) const override {
+        if (m_pos == INVALID_POS) return false;
+        auto cur = coord();
+        auto next = positions[pos];
+        return cur[R] - cur[F] == next[R] - next[F] || cur[R] + cur[F] == next[R] + next[F];
+    }
+
+    virtual void reinit(const State& state) override;
+};
+
+struct Rook : public Piece {
+    Rook(Colors color, char pos) : Piece(Types::Rook, color, pos, {}) {}
+
+    virtual bool movePattern(char pos) const override {
+        if (m_pos == INVALID_POS) return false;
+        auto cur = coord();
+        auto next = positions[pos];
+        return cur[R] == next[R] || cur[F] == next[F];
+    }
+
+    virtual void reinit(const State& state) override;
+};
+
+struct Queen : public Piece {
+    Queen(Colors color, char pos) : Piece(Types::Queen, color, pos, {}) {}
+
+    virtual bool movePattern(char pos) const override {
+        if (m_pos == INVALID_POS) return false;
+        auto cur = coord();
+        auto next = positions[pos];
+        return cur[R] == next[R] || cur[F] == next[F] || cur[R] - cur[F] == next[R] - next[F] || cur[R] + cur[F] == next[R] + next[F];
+    }
+
+    virtual void reinit(const State& state) override;
+};
+
+struct King : public Piece {
+    King(Colors color, char pos) : Piece(Types::King, color, pos, {}) {}
+
+    virtual bool movePattern(char pos) const override {
+        if (m_pos == INVALID_POS) return false;
+        auto cur = coord();
+        auto next = positions[pos];
+        Direction diff = {char(next[R] - cur[R]), char(next[F] - cur[F])};
+        return diff[R]*diff[R] + diff[F]*diff[F] <= 2;
+    }
+
+    virtual void reinit(const State& state) override;
+};
+
+struct PieceSet {
+    Piece* begin() { return &p1; }
+    Piece* end() { return &r2 + 1; }
+    const Piece* begin() const { return &p1; }
+    const Piece* end() const { return &r2 + 1; }
+    Piece& operator[](int i) { return *(begin() + i); }
+    const Piece& operator[](int i) const { return *(begin() + i); }
+
+    Pawn   p1;
+    Pawn   p2;
+    Pawn   p3;
+    Pawn   p4;
+    Pawn   p5;
+    Pawn   p6;
+    Pawn   p7;
+    Pawn   p8;
+    Rook   r1;
+    Knight n1;
+    Bishop b1;
+    Queen  q;
+    King   k;
+    Bishop b2;
+    Knight n2;
+    Rook   r2;
+};
+
+struct State {
+    State();
+    PieceSet blacks;
+    PieceSet whites;
+    Board board;
+    Pins blackPins;
+    Pins whitePins;
+};
+
+Direction findPin(const Piece& piece, const State& state) {
+    auto& pins = piece.color() ? state.blackPins : state.whitePins;
+    auto it = std::find_if(pins.begin(), pins.end(), [&] (const Pin& pin) { return pin.pinned == &piece; });
+    if (it != pins.end()) return it->d;
+    return {0, 0};
+}
+
+struct Find {
+    Find(const Board& board) : m_board(board) {}
+    bool operator() (char pos) const { return m_board[pos]; }
+    const Board& m_board;
+};
+
+struct Add {
+    Add(const Board& board, std::set<char>& moves, Piece::Colors color) : m_board(board), m_moves(moves), m_color(color) {}
+    bool operator() (char pos) const {
+        if (!m_board[pos] || m_board[pos]->color() != m_color) m_moves.insert(pos);
+        return m_board[pos];
+    }
+    const Board& m_board;
+    std::set<char>& m_moves;
+    Piece::Colors m_color;
+};
+
+void Pawn::reinit(const State& state) {
+    if (m_pos == INVALID_POS) return;
+    if (!m_update) return;
+    m_update = false;
+    m_allowed.clear();
+
+    auto pin = findPin(*this, state);
+
+    auto & left = m_color ? SW : NW;
+    auto & right = m_color ? SE : NE;
+
+    for (auto& direction : filter(pin, { left, right })) {
+        auto pos = makeStep(m_pos, direction);
+        if (pos != INVALID_POS && state.board[pos] && state.board[pos]->color() != m_color) m_allowed.insert(pos);
+    }
+
+    auto & forward = m_color ? S : N;
+    if (!filter(pin, {forward}).empty()) {
+        traverse(m_pos, forward, [&] (char pos) {
+                if (!state.board[pos]) m_allowed.insert(pos);
+                return state.board[pos] || !is_first_move();
+            }, 2);
+    }
+}
+
+void Knight::reinit(const State& state) {
+    if (m_pos == INVALID_POS) return;
+    if (!m_update) return;
+    m_update = false;
+    m_allowed.clear();
+    auto pin = findPin(*this, state);
+    if (pin[R] != 0 || pin[F] != 0) return;
+    for (auto& direction : { NNE, ENE, ESE, SSE, SSW, WSW, WNW, NNW }) {
+        auto pos = makeStep(m_pos, direction);
+        if (pos != INVALID_POS && (!state.board[pos] || state.board[pos]->color() != m_color)) m_allowed.insert(pos);
+    }
+}
+
+void Bishop::reinit(const State& state) {
+    if (m_pos == INVALID_POS) return;
+    if (!m_update) return;
+    m_update = false;
+    m_allowed.clear();
+    auto pin = findPin(*this, state);
+    for (auto& direction : filter(pin, { NE, SE, SW, NW })) {
+        traverse(m_pos, direction, Add(state.board, m_allowed, m_color));
+    }
+}
+
+void Rook::reinit(const State& state) {
+    if (m_pos == INVALID_POS) return;
+    if (!m_update) return;
+    m_update = false;
+    m_allowed.clear();
+    auto pin = findPin(*this, state);
+    for (auto& direction : filter(pin, { N, E, S, W })) {
+        traverse(m_pos, direction, Add(state.board, m_allowed, m_color));
+    }
+}
+
+void Queen::reinit(const State& state) {
+    if (m_pos == INVALID_POS) return;
+    if (!m_update) return;
+    m_update = false;
+    m_allowed.clear();
+    auto pin = findPin(*this, state);
+    for (auto& direction : filter(pin, { N, NE, E, SE, S, SW, W, NW })) {
+        traverse(m_pos, direction, Add(state.board, m_allowed, m_color));
+    }
+}
+
+void King::reinit(const State& state) {
+    if (m_pos == INVALID_POS) return;
+    if (!m_update) return;
+    m_update = false;
+    m_allowed.clear();
+    auto& enemyPieces = m_color ? state.whites : state.blacks;
+    auto& pawnAttackLeft = m_color ? SW : NW;
+    auto& pawnAttackRight = m_color ? SE : NE;
+    for (auto& direction : { N, NE, E, SE, S, SW, W, NW }) {
+        auto pos = makeStep(m_pos, direction);
+        bool accept = pos != INVALID_POS && !(state.board[pos] && state.board[pos]->color() == m_color);
+        if (accept) {
+            for (auto& p : enemyPieces) {
+                if (!p.movePattern(pos)) continue;
+                if (p.type() == Piece::Knight || p.type() == Piece::King) {
+                    accept = false;
+                    break;
+                }
+                else if (p.type() == Piece::Pawn) {
+                    auto from = positions[pos];
+                    auto to = p.coord();
+                    Direction d {char(to[R] - from[R]), char(to[F] - from[F])};
+                    if (d == pawnAttackLeft || d == pawnAttackRight) {
+                        accept = false;
+                        break;
+                    }
+                }
+                else {
+                    auto from = positions[pos];
+                    auto to = p.coord();
+                    Direction d = normalize({char(to[R] - from[R]), char(to[F] - from[F])});
+                    auto reached = traverse(pos, d, Find(state.board));
+                    if (p.pos() == reached) {
+                        accept = false;
+                        break;
+                    }
+                }
+            }
+        }
+        if (accept) m_allowed.insert(pos);
+    }
+}
+
+const char* Piece::name() const {
+    static_assert(pieceNames.size() == Piece::NUM_PIECES, "Mismatch between piece names and types");
+    return pieceNames[m_type];
+}
+
+char Piece::initial() const {
+    static_assert(blackShort.size() == Piece::NUM_PIECES, "Mismatch between piece names and types");
+    static_assert(whiteShort.size() == Piece::NUM_PIECES, "Mismatch between piece names and types");
+    return m_color ? blackShort[m_type] : whiteShort[m_type];
+}
+
+void Piece::invalidate() {
+    m_update = true;
+}
+
+
+const char* Piece::coord() const {
+    if (m_pos == INVALID_POS) return "";
+    return positions[m_pos];
+}
+
+bool Piece::canReach(char pos) const {
+    return movePattern(pos) && m_allowed.count(pos);
+}
+
+void Piece::take() {
+    m_pos = INVALID_POS;
+    m_allowed = {};
+}
+
+State::State()
+    : blacks {
+        {Piece::Black, "a7"_P, {"a5"_P, "a6"_P} },
+        {Piece::Black, "b7"_P, {"b5"_P, "b6"_P} },
+        {Piece::Black, "c7"_P, {"c5"_P, "c6"_P} },
+        {Piece::Black, "d7"_P, {"d5"_P, "d6"_P} },
+        {Piece::Black, "e7"_P, {"e5"_P, "e6"_P} },
+        {Piece::Black, "f7"_P, {"f5"_P, "f6"_P} },
+        {Piece::Black, "g7"_P, {"g5"_P, "g6"_P} },
+        {Piece::Black, "h7"_P, {"h5"_P, "h6"_P} },
+        {Piece::Black, "a8"_P},
+        {Piece::Black, "b8"_P, {"a6"_P, "c6"_P} },
+        {Piece::Black, "c8"_P},
+        {Piece::Black, "d8"_P},
+        {Piece::Black, "e8"_P},
+        {Piece::Black, "f8"_P},
+        {Piece::Black, "g8"_P, {"f6"_P, "h6"_P} },
+        {Piece::Black, "h8"_P},
+    }
+    , whites {
+        {Piece::White, "a2"_P, {"a3"_P, "a4"_P} },
+        {Piece::White, "b2"_P, {"b3"_P, "b4"_P} },
+        {Piece::White, "c2"_P, {"c3"_P, "c4"_P} },
+        {Piece::White, "d2"_P, {"d3"_P, "d4"_P} },
+        {Piece::White, "e2"_P, {"e3"_P, "e4"_P} },
+        {Piece::White, "f2"_P, {"f3"_P, "f4"_P} },
+        {Piece::White, "g2"_P, {"g3"_P, "g4"_P} },
+        {Piece::White, "h2"_P, {"h3"_P, "h4"_P} },
+        {Piece::White, "a1"_P},
+        {Piece::White, "b1"_P, {"a3"_P, "c3"_P} },
+        {Piece::White, "c1"_P},
+        {Piece::White, "d1"_P},
+        {Piece::White, "e1"_P},
+        {Piece::White, "f1"_P},
+        {Piece::White, "g1"_P, {"f3"_P, "h3"_P} },
+        {Piece::White, "h1"_P},
+    }
+    , board {{
+        &whites[ 8],  &whites[ 9],  &whites[10],  &whites[11],  &whites[12],  &whites[13],  &whites[14],  &whites[15],
+        &whites[ 0],  &whites[ 1],  &whites[ 2],  &whites[ 3],  &whites[ 4],  &whites[ 5],  &whites[ 6],  &whites[ 7],
+        nullptr,      nullptr,      nullptr,      nullptr,      nullptr,      nullptr,      nullptr,      nullptr,
+        nullptr,      nullptr,      nullptr,      nullptr,      nullptr,      nullptr,      nullptr,      nullptr,
+        nullptr,      nullptr,      nullptr,      nullptr,      nullptr,      nullptr,      nullptr,      nullptr,
+        nullptr,      nullptr,      nullptr,      nullptr,      nullptr,      nullptr,      nullptr,      nullptr,
+        &blacks[ 0],  &blacks[ 1],  &blacks[ 2],  &blacks[ 3],  &blacks[ 4],  &blacks[ 5],  &blacks[ 6],  &blacks[ 7],
+        &blacks[ 8],  &blacks[ 9],  &blacks[10],  &blacks[11],  &blacks[12],  &blacks[13],  &blacks[14],  &blacks[15],
+    }}
+{}
+
+Chessboard::Chessboard()
+    : m_state(new State())
+{
+    setGrammar();
+}
+
+Chessboard::~Chessboard() = default;
+
+void Chessboard::setPrompt(const std::string& prompt) {
+    m_prompt = prompt;
+    setGrammar();
+}
+
+void Chessboard::setGrammar() {
+    m_grammar.clear();
+
+    std::string result;
+    if (m_prompt.empty()) {
+        result += "move ::= \" \" ((piece | frompos) \" \" \"to \"?)? topos\n";
+        //result += "move ::= \" \" frompos \" \" \"to \"? topos\n";
+    }
+    else {
+        // result += "move ::= prompt \" \" ((piece | frompos) \" \" \"to \"?)? topos\n"
+        result += "move ::= prompt \" \" frompos \" \" \"to \"? topos\n"
+        "prompt ::= \" " + m_prompt + "\"\n";
+    }
+
+    std::set<Piece::Types> pieceTypes;
+    std::set<char> from_pos;
+    std::set<char> to_pos;
+    auto& pieces =  m_moveCounter % 2 ? m_state->blacks : m_state->whites;
+    std::set<size_t> flags;
+    for (auto& p : pieces) {
+        if (p.allowed().empty()) continue;
+        bool addPiece = false;
+        if (!m_inCheck || p.type() == Piece::King) {
+            to_pos.insert(p.allowed().begin(), p.allowed().end());
+            addPiece = !p.allowed().empty();
+        }
+        else {
+            for (auto move : p.allowed()) {
+                if (m_allowedInCheck.count(move)) {
+                    to_pos.insert(move);
+                    addPiece = true;
+                }
+            }
+        }
+        if (addPiece) {
+            pieceTypes.insert(p.type());
+            from_pos.insert(p.pos());
+        }
+    }
+    if (pieceTypes.empty()) return;
+
+    result += "piece ::= (";
+    for (auto& p : pieceTypes) result += " \"" + std::string(pieceNames[p]) + "\" |";
+    result.pop_back();
+    result += ")\n\n";
+
+    result += "frompos ::= (";
+    for (auto& p : from_pos) result += " \"" + std::string(positions[p]) + "\" |";
+    result.pop_back();
+    result += ")\n";
+
+    result += "topos ::= (";
+    for (auto& p : to_pos) result += " \"" + std::string(positions[p]) + "\" |";
+    result.pop_back();
+    result += ")\n";
+
+    m_grammar = std::move(result);
+}
+
+std::string Chessboard::stringifyBoard() {
+    std::string result;
+    result.reserve(16 + 2 * 64 + 16);
+    for (char rank = 'a'; rank <= 'h'; ++rank) {
+        result.push_back(rank);
+        result.push_back(' ');
+    }
+    result.back() = '\n';
+    for (int i = 7; i >= 0; --i) {
+        for (int j = 0; j < 8; ++j) {
+            auto p = m_state->board[i * 8 + j];
+            if (p) result.push_back(p->initial());
+            else result.push_back((i + j) % 2 ? '.' : '*');
+            result.push_back(' ');
+        }
+        result.push_back('0' + i + 1);
+        result.push_back('\n');
+    }
+    return result;
+}
+
+std::string Chessboard::process(const std::string& command) {
+    const auto t_start = std::chrono::high_resolution_clock::now();
+    auto color = Piece::Colors(m_moveCounter % 2);
+    Piece* piece = nullptr;
+    auto pos_to = INVALID_POS;
+    if (!parseCommand(command, piece, pos_to)) return "";
+
+    auto pos_from = piece->pos();
+
+    if (!move(*piece, pos_to)) return "";
+
+    flagUpdates(pos_from, pos_to);
+
+    detectChecks();
+
+    auto& enemyPieces = color ? m_state->whites : m_state->blacks;
+    for (auto& p : enemyPieces) p.reinit(*m_state); // only enemy moves needed next
+
+    std::string result = {positions[pos_from][R], positions[pos_from][F], '-', positions[pos_to][R], positions[pos_to][F]};
+    ++m_moveCounter;
+    setGrammar();
+    const auto t_end = std::chrono::high_resolution_clock::now();
+    auto t_ms = std::chrono::duration_cast<std::chrono::milliseconds>(t_end - t_start).count();
+    fprintf(stdout, "%s: Move '%s%s%s', (t = %d ms)\n", __func__, "\033[1m", result.data(), "\033[0m", (int) t_ms);
+    if (m_grammar.empty()) result.push_back('#');
+    return result;
+}
+
+bool Chessboard::parseCommand(const std::string& command, Piece*& piece, char& pos_to) {
+    auto color = Piece::Colors(m_moveCounter % 2);
+    fprintf(stdout, "%s: Command to %s: '%s%.*s%s'\n", __func__, (color ? "Black" : "White"), "\033[1m", int(command.size()), command.data(), "\033[0m");
+
+    if (command.empty()) return false;
+    auto tokens = split(command, ' ');
+    auto pos_from = INVALID_POS;
+    auto type = Piece::Types::NUM_PIECES;
+    if (tokens.size() == 1) {
+        type = Piece::Types::Pawn;
+        pos_to = strToPos(tokens.front());
+    }
+    else {
+        pos_from = strToPos(tokens.front());
+        if (pos_from == INVALID_POS) type = Piece::Types(strToType(tokens.front()));
+        pos_to = strToPos(tokens.back());
+    }
+    if (pos_to == INVALID_POS) return false;
+    if (pos_from == INVALID_POS) {
+        if (type == Piece::Types::NUM_PIECES) return false;
+        auto& pieces = color ? m_state->blacks : m_state->whites;
+        for (auto& p : pieces) {
+            if (p.type() == type && p.canReach(pos_to)) {
+                pos_from = p.pos();
+                break;
+            }
+        }
+    }
+    if (pos_from == INVALID_POS) return false;
+    if (m_state->board[pos_from] == nullptr) return false;
+    piece = m_state->board[pos_from];
+    if (piece->color() != color) return false;
+    return true;
+}
+
+void Chessboard::flagUpdates(char pos_from, char pos_to) {
+    auto color = Piece::Colors(m_moveCounter % 2);
+    auto& enemyPieces = color ? m_state->whites : m_state->blacks;
+    auto& ownPieces = color ? m_state->blacks : m_state->whites;
+    for (auto& p : enemyPieces) {
+        if (p.movePattern(pos_to) || p.movePattern(pos_from)) {
+            updatePins(p);
+            p.invalidate();
+        }
+    }
+
+    for (auto& p : ownPieces) {
+        if (p.movePattern(pos_to) || p.movePattern(pos_from)) {
+            updatePins(p);
+            p.invalidate();
+        }
+    }
+}
+
+void Chessboard::updatePins(Piece& piece) {
+    if (piece.type() == Piece::Pawn || piece.type() == Piece::Knight || piece.type() == Piece::King) return;
+    auto& enemyPieces = piece.color() ? m_state->whites : m_state->blacks;
+    auto& enemyPins = piece.color() ? m_state->whitePins : m_state->blackPins;
+    auto& king = enemyPieces.k;
+    auto it = std::find_if(enemyPins.begin(), enemyPins.end(), [&] (const Pin& pin) { return pin.pinner == &piece; });
+    if (it != enemyPins.end()) {
+        it->pinned->invalidate();
+        enemyPins.erase(it);
+    }
+    if (piece.movePattern(king.pos())) {
+        auto to = positions[king.pos()];
+        auto from = piece.coord();
+        Direction d = normalize({char(to[R] - from[R]), char(to[F] - from[F])});
+
+        auto reached = traverse(piece.pos(), d, Find(m_state->board));
+        auto foundPiece = m_state->board[reached];
+        if (&king == foundPiece) {
+            // check
+            king.invalidate();
+        }
+        else if (foundPiece && foundPiece->color() != piece.color()) {
+            reached = traverse(reached, d, Find(m_state->board));
+            if (&king == m_state->board[reached]) {
+                enemyPins.push_back({d, &piece, foundPiece});
+                foundPiece->invalidate();
+            }
+        }
+    }
+}
+
+void Chessboard::detectChecks() {
+    auto color = Piece::Colors(m_moveCounter % 2);
+    auto& enemyPieces = color ? m_state->whites : m_state->blacks;
+    auto& ownPieces = color ? m_state->blacks : m_state->whites;
+    auto& king = enemyPieces.k;
+    auto& pawnAttackLeft = color ? SW : NW;
+    auto& pawnAttackRight = color ? SE : NE;
+    for (auto& p : ownPieces) {
+        if (!p.movePattern(king.pos())) continue;
+        auto to = positions[king.pos()];
+        auto from = p.coord();
+
+        if (p.type() == Piece::Knight) {
+            if (!m_inCheck) {
+                m_allowedInCheck = { p.pos() };
+            }
+            else {
+                m_allowedInCheck.clear();
+            }
+            m_inCheck = true;
+        }
+        else if (p.type() == Piece::Pawn) {
+            Direction d {char(to[R] - from[R]), char(to[F] - from[F])};
+            if (d == pawnAttackLeft || d == pawnAttackRight) {
+                if (!m_inCheck) {
+                    m_allowedInCheck = { p.pos() };
+                }
+                else {
+                    m_allowedInCheck.clear();
+                }
+                m_inCheck = true;
+            }
+        }
+        else {
+            Direction d = normalize({char(to[R] - from[R]), char(to[F] - from[F])});
+            std::set<char> tmp;
+            auto pos = traverse(p.pos(), d, Add(m_state->board, tmp, king.color()));
+            if (pos == king.pos()) {
+                tmp.insert(p.pos());
+                if (!m_inCheck) {
+                    m_allowedInCheck = std::move(tmp);
+                }
+                else {
+                    m_allowedInCheck.clear();
+                }
+                m_inCheck = true;
+            }
+        }
+    }
+}
+
+bool Chessboard::move(Piece& piece, char pos_to) {
+    auto& allowed = piece.allowed();
+
+    if (allowed.count(pos_to) == 0 || (m_inCheck && piece.type() != Piece::King && m_allowedInCheck.count(pos_to) == 0)) return false;
+    if (m_state->board[pos_to] && m_state->board[pos_to]->color() == piece.color()) return false;
+    if (m_state->board[pos_to]) m_state->board[pos_to]->take();
+    m_state->board[piece.pos()] = nullptr;
+    m_state->board[pos_to] = &piece;
+    piece.setPos(pos_to);
+
+    m_inCheck = false;
+    m_allowedInCheck.clear();
+
+    return true;
+}

examples/wchess/libwchess/Chessboard.h

@@ -0,0 +1,33 @@
+#pragma once
+#include <string>
+#include <set>
+#include <memory>
+
+// just basic validation
+// fixme: missing en passant, castling, promotion, etc.
+struct State;
+class Piece;
+class Chessboard {
+public:
+    Chessboard();
+    ~Chessboard();
+    std::string process(const std::string& command);
+    std::string stringifyBoard();
+    const std::string& grammar() { return m_grammar; }
+    const std::string& prompt() { return m_prompt; }
+    void setPrompt(const std::string& prompt);
+private:
+    bool parseCommand(const std::string& command, Piece*& piece, char& pos_to);
+    bool move(Piece& piece, char pos);
+    void flagUpdates(char pos_from, char pos_to);
+    void updatePins(Piece& piece);
+    void detectChecks();
+    void setGrammar();
+
+    std::unique_ptr<State> m_state;
+    std::set<char> m_allowedInCheck;
+    bool m_inCheck = false;
+    int m_moveCounter = 0;
+    std::string m_grammar;
+    std::string m_prompt;
+};

examples/wchess/libwchess/WChess.cpp

@@ -0,0 +1,193 @@
+#include "WChess.h"
+#include "Chessboard.h"
+#include "grammar-parser.h"
+#include "common.h"
+#include <thread>
+
+WChess::WChess(whisper_context * ctx,
+        const whisper_full_params & wparams,
+        callbacks cb,
+        settings s)
+        : m_ctx(ctx)
+        , m_wparams(wparams)
+        , m_cb(cb)
+        , m_settings(s)
+        , m_board(new Chessboard())
+{}
+
+WChess::~WChess() = default;
+
+void WChess::set_move(const std::string& moves, float prob) const {
+    if (m_cb.set_move) (*m_cb.set_move)(moves, prob);
+}
+
+void WChess::set_grammar(const std::string& grammar) const {
+    if (m_cb.set_grammar) (*m_cb.set_grammar)(grammar);
+}
+
+bool WChess::get_audio(std::vector<float>& pcmf32) const {
+    if (m_cb.get_audio) return (*m_cb.get_audio)(pcmf32);
+    return false;
+}
+
+std::string WChess::stringify_board() const {
+    return m_board->stringifyBoard();
+}
+
+std::string WChess::get_grammar() const {
+    return m_board->grammar();
+}
+
+void WChess::run() {
+    bool have_prompt  = true;
+    bool ask_prompt   = !have_prompt;
+
+    float logprob_min  = 0.0f;
+
+    float logprob_sum  = 0.0f;
+
+    int n_tokens  = 0;
+
+    std::vector<float> pcmf32_cur;
+    std::vector<float> pcmf32_prompt;
+
+    const std::string k_prompt = have_prompt ? "" : "rook to d4, f3";
+    int64_t t_ms = 0;
+
+    if (ask_prompt) {
+        fprintf(stdout, "\n");
+        fprintf(stdout, "%s: Say the following phrase: '%s%s%s'\n", __func__, "\033[1m", k_prompt.c_str(), "\033[0m");
+        fprintf(stdout, "\n");
+
+        ask_prompt = false;
+    }
+
+    while (get_audio(pcmf32_cur)) {
+        if (!pcmf32_cur.empty()) {
+            // fprintf(stdout, "%s: Processing ...\n", __func__);
+
+            if (!have_prompt) {
+                const auto txt = ::trim(transcribe(pcmf32_cur, logprob_min, logprob_sum, n_tokens, t_ms));
+
+                fprintf(stdout, "%s: Heard '%s%s%s', (t = %d ms)\n", __func__, "\033[1m", txt.c_str(), "\033[0m", (int) t_ms);
+
+                const float sim = similarity(txt, k_prompt);
+
+                if (txt.length() < 0.8*k_prompt.length() || txt.length() > 1.2*k_prompt.length() || sim < 0.8f) {
+                    fprintf(stdout, "%s: WARNING: prompt not recognized, try again\n", __func__);
+                    ask_prompt = true;
+                } else {
+                    fprintf(stdout, "\n");
+                    fprintf(stdout, "%s: The prompt has been recognized!\n", __func__);
+                    fprintf(stdout, "%s: Waiting for voice commands ...\n", __func__);
+                    fprintf(stdout, "\n");
+
+                    // save the audio for the prompt
+                    pcmf32_prompt = pcmf32_cur;
+                    have_prompt = true;
+                    m_board->setPrompt(k_prompt);
+                }
+            } else {
+                if (!pcmf32_prompt.empty()) pcmf32_cur.insert(pcmf32_cur.begin(), pcmf32_prompt.begin(), pcmf32_prompt.end());
+                constexpr size_t MIN_SIZE = 1.2 * WHISPER_SAMPLE_RATE;
+                if (MIN_SIZE > pcmf32_cur.size()) pcmf32_cur.insert(pcmf32_cur.begin(), MIN_SIZE - pcmf32_cur.size(), 0.0f);
+
+                // fprintf(stdout, "%s: grammar rules:\n'%s'\n", __func__, m_board->grammar().c_str());
+
+                auto grammar_parsed = grammar_parser::parse(m_board->grammar().c_str());
+                auto grammar_rules  = grammar_parsed.c_rules();
+
+                m_wparams.grammar_rules   = grammar_rules.data();
+                m_wparams.n_grammar_rules = grammar_rules.size();
+
+                m_wparams.i_start_rule    = grammar_parsed.symbol_ids.at("move");
+                auto txt = ::trim(transcribe(pcmf32_cur, logprob_min, logprob_sum, n_tokens, t_ms));
+
+                const float p = 100.0f * std::exp(logprob_min);
+
+                fprintf(stdout, "%s: heard '%s'\n", __func__, txt.c_str());
+
+                // find the prompt in the text
+                float best_sim = 0.0f;
+                size_t best_len = 0;
+                for (int n = 0.8*k_prompt.size(); n <= 1.2*k_prompt.size(); ++n) {
+                    const auto prompt = txt.substr(0, n);
+
+                    const float sim = similarity(prompt, k_prompt);
+
+                    //fprintf(stderr, "%s: prompt = '%s', sim = %f\n", __func__, prompt.c_str(), sim);
+
+                    if (sim > best_sim) {
+                        best_sim = sim;
+                        best_len = n;
+                    }
+                }
+
+                fprintf(stdout, "%s:   DEBUG: txt = '%s', prob = %.2f%%\n", __func__, txt.c_str(), p);
+                std::string command = ::trim(txt.substr(best_len));
+
+                fprintf(stdout, "%s: Command '%s%s%s', (t = %d ms)\n", __func__, "\033[1m", command.c_str(), "\033[0m", (int) t_ms);
+                fprintf(stdout, "\n");
+
+                if (!command.empty()) {
+                    set_move(m_board->process(command), p);
+                    set_grammar(m_board->grammar());
+                }
+                if (m_board->grammar().empty()) {
+                    fprintf(stdout, "%s: No more moves possible\n", __func__);
+                    break;
+                }
+            }
+        }
+
+        if (ask_prompt) {
+            fprintf(stdout, "\n");
+            fprintf(stdout, "%s: Say the following phrase: '%s%s%s'\n", __func__, "\033[1m", k_prompt.c_str(), "\033[0m");
+            fprintf(stdout, "\n");
+
+            ask_prompt = false;
+        }
+    }
+}
+
+std::string WChess::transcribe(
+                const std::vector<float> & pcmf32,
+                float & logprob_min,
+                float & logprob_sum,
+                int & n_tokens,
+                int64_t & t_ms) {
+    const auto t_start = std::chrono::high_resolution_clock::now();
+
+    logprob_min = 0.0f;
+    logprob_sum = 0.0f;
+    n_tokens    = 0;
+    t_ms = 0;
+
+    if (whisper_full(m_ctx, m_wparams, pcmf32.data(), pcmf32.size()) != 0) {
+        return {};
+    }
+
+    std::string result;
+
+    const int n_segments = whisper_full_n_segments(m_ctx);
+    for (int i = 0; i < n_segments; ++i) {
+        const char * text = whisper_full_get_segment_text(m_ctx, i);
+
+        result += text;
+
+        const int n = whisper_full_n_tokens(m_ctx, i);
+        for (int j = 0; j < n; ++j) {
+            const auto token = whisper_full_get_token_data(m_ctx, i, j);
+
+            if(token.plog > 0.0f) return {};
+            logprob_min = std::min(logprob_min, token.plog);
+            logprob_sum += token.plog;
+            ++n_tokens;
+        }
+    }
+
+    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;
+}

examples/wchess/libwchess/WChess.h

@@ -0,0 +1,63 @@
+#pragma once
+#include "whisper.h"
+#include <string>
+#include <vector>
+#include <memory>
+
+class Chessboard;
+
+class WChess {
+public:
+    using CheckRunningCb = bool (*)();
+    using GetAudioCb = bool (*)(std::vector<float> &);
+    using SetMovesCb = void (*)(const std::string &, float);
+    using SetGrammarCb = void (*)(const std::string &);
+    using ClearAudioCb = void (*)();
+
+    struct callbacks {
+        GetAudioCb get_audio = nullptr;
+        SetMovesCb set_move = nullptr;
+        SetGrammarCb set_grammar = nullptr;
+    };
+
+    struct settings {
+        int32_t vad_ms     = 2000;
+        int32_t prompt_ms  = 5000;
+        int32_t command_ms = 4000;
+        float vad_thold    = 0.2f;
+        float freq_thold   = 100.0f;
+        bool print_energy  = false;
+    };
+
+    WChess(
+        whisper_context * ctx,
+        const whisper_full_params & wparams,
+        callbacks cb,
+        settings s
+    );
+    ~WChess();
+
+    void run();
+
+    std::string stringify_board() const;
+
+    std::string get_grammar() const;
+
+private:
+    bool get_audio(std::vector<float>& pcmf32) const;
+    void set_move(const std::string& moves, float prob) const;
+    void set_grammar(const std::string& grammar) const;
+
+    std::string transcribe(
+                    const std::vector<float> & pcmf32,
+                    float & logprob_min,
+                    float & logprob_sum,
+                    int & n_tokens,
+                    int64_t & t_ms);
+
+    whisper_context * m_ctx;
+    whisper_full_params m_wparams;
+    const callbacks m_cb;
+    const settings m_settings;
+    std::unique_ptr<Chessboard> m_board;
+};

examples/wchess/libwchess/test-chessboard.cpp

@@ -0,0 +1,117 @@
+#include "Chessboard.h"
+
+#define ASSERT(x) \
+    do { \
+        if (!(x)) { \
+            fprintf(stderr, "ASSERT: %s:%d: %s\n", __FILE__, __LINE__, #x); \
+            fflush(stderr); \
+            exit(1); \
+        } \
+    } while (0)
+
+
+int main() {
+    {
+        Chessboard chess;
+
+        ASSERT(chess.process("pawn to d4") == "d2-d4");
+        ASSERT(chess.process("e5") == "e7-e5");
+        ASSERT(chess.process("c1 h6") == "c1-h6");
+        ASSERT(chess.process("queen h4") == "d8-h4");
+        ASSERT(chess.process("bishop to g5") == "h6-g5");
+        ASSERT(chess.process("bishop to b4") == "f8-b4");
+        ASSERT(chess.process("c4") == "");
+        ASSERT(chess.process("knight c3") == "b1-c3");
+        ASSERT(chess.process("knight c6") == "b8-c6");
+        ASSERT(chess.process("f3") == "");
+    }
+
+    {
+        Chessboard chess;
+
+        ASSERT(chess.process("d4") == "d2-d4");
+        ASSERT(chess.process("e5") == "e7-e5");
+        ASSERT(chess.process("e4") == "e2-e4");
+        ASSERT(chess.process("queen h4") == "d8-h4");
+        ASSERT(chess.process("queen h5") == "d1-h5");
+        ASSERT(chess.process("f5") == "");
+        ASSERT(chess.process("g6") == "g7-g6");
+        ASSERT(chess.process("knight e2") == "g1-e2");
+        ASSERT(chess.process("f5") == "f7-f5");
+        ASSERT(chess.process("knight g3") == "e2-g3");
+        ASSERT(chess.process("g5") == "");
+        ASSERT(chess.process("king e7") == "e8-e7");
+        ASSERT(chess.process("f4") == "f2-f4");
+        ASSERT(chess.process("g5") == "g6-g5");
+    }
+
+    {
+        Chessboard chess;
+
+        ASSERT(chess.process("e4") == "e2-e4");
+        ASSERT(chess.process("c5") == "c7-c5");
+        ASSERT(chess.process("e5") == "e4-e5");
+        ASSERT(chess.process("c4") == "c5-c4");
+        ASSERT(chess.process("e6") == "e5-e6");
+        ASSERT(chess.process("c3") == "c4-c3");
+        ASSERT(chess.process("e7") == "");
+        ASSERT(chess.process("f7") == "e6-f7");
+        ASSERT(chess.process("d2") == "");
+        ASSERT(chess.process("king to f7") == "e8-f7");
+        ASSERT(chess.process("f4") == "f2-f4");
+        ASSERT(chess.process("d2") == "c3-d2");
+        ASSERT(chess.process("f5") == "");
+        ASSERT(chess.process("king to e2") == "e1-e2");
+        ASSERT(chess.process("king to g6") == "f7-g6");
+        ASSERT(chess.process("f5") == "f4-f5");
+        ASSERT(chess.process("e6") == "");
+        ASSERT(chess.process("king to h5") == "g6-h5");
+        ASSERT(chess.process("g4") == "g2-g4");
+        ASSERT(chess.process("king to g5") == "h5-g5");
+        ASSERT(chess.process("h4") == "h2-h4");
+        ASSERT(chess.process("king to h5") == "");
+        ASSERT(chess.process("king to g6") == "");
+        ASSERT(chess.process("king to h6") == "g5-h6");
+        ASSERT(chess.process("bishop to d2") == "c1-d2");
+        ASSERT(chess.process("king to g5") == "");
+        ASSERT(chess.process("g5") == "g7-g5");
+    }
+
+    {
+        Chessboard chess;
+        ASSERT(chess.process("f4") == "f2-f4");
+        ASSERT(chess.process("e5") == "e7-e5");
+        ASSERT(chess.process("g4") == "g2-g4");
+        ASSERT(chess.process("queen to h4") == "d8-h4#");
+        ASSERT(chess.process("knight f3") == "");
+        ASSERT(chess.grammar().empty());
+    }
+
+    {
+        Chessboard chess;
+        ASSERT(chess.process("f4") == "f2-f4");
+        ASSERT(chess.process("e5") == "e7-e5");
+        ASSERT(chess.process("g4") == "g2-g4");
+        ASSERT(chess.process("d5") == "d7-d5");
+        ASSERT(chess.process("g1 f3") == "g1-f3");
+        ASSERT(chess.process("queen to h4") == "d8-h4");
+        ASSERT(!chess.grammar().empty());
+    }
+
+    {
+        Chessboard chess;
+        ASSERT(chess.process("knight c3") == "b1-c3");
+        ASSERT(chess.process("knight c6") == "b8-c6");
+        ASSERT(chess.process("knight b5") == "c3-b5");
+        ASSERT(chess.process("knight f6") == "g8-f6");
+        ASSERT(chess.process("knight d6") == "b5-d6");
+        ASSERT(chess.process("knight d4") == "");
+        ASSERT(chess.process("d6") == "c7-d6");
+        ASSERT(chess.process("e4") == "e2-e4");
+        ASSERT(chess.process("knight d4") == "c6-d4");
+        ASSERT(chess.process("d3") == "d2-d3");
+        ASSERT(chess.process("knight e4") == "f6-e4");
+        ASSERT(chess.process("king to e2") == "");
+        ASSERT(chess.process("king to d2") == "");
+    }
+}

examples/wchess/wchess.cmd/CMakeLists.txt

@@ -0,0 +1,8 @@
+if (WHISPER_SDL2)
+    set(TARGET wchess)
+    add_executable(${TARGET} wchess.cmd.cpp)
+
+    include(DefaultTargetOptions)
+
+    target_link_libraries(${TARGET} PRIVATE wchess-core common-sdl ${CMAKE_THREAD_LIBS_INIT})
+endif ()

examples/wchess/wchess.cmd/wchess.cmd.cpp

@@ -0,0 +1,247 @@
+// Command line voice assisted chess
+//
+// Speak chess move commands to the microphone.
+// The moves will translated to chessboard positions.
+//
+//
+
+#include "WChess.h"
+#include "common-sdl.h"
+#include <iostream>
+
+#include <memory>
+#include <thread>
+
+// command-line parameters
+struct whisper_params {
+    int32_t n_threads  = std::min(4, (int32_t) std::thread::hardware_concurrency());
+    int32_t prompt_ms  = 5000;
+    int32_t command_ms = 8000;
+    int32_t capture_id = -1;
+    int32_t max_tokens = 32;
+    int32_t audio_ctx  = 0;
+
+    float vad_thold  = 0.6f;
+    float freq_thold = 100.0f;
+
+    float grammar_penalty = 100.0f;
+
+    bool speed_up      = false;
+    bool translate     = false;
+    bool print_special = false;
+    bool print_energy  = false;
+    bool no_timestamps = true;
+    bool use_gpu       = true;
+
+    std::string language  = "en";
+    std::string model     = "models/ggml-base.en.bin";
+    std::string fname_out;
+    std::string commands;
+    std::string prompt;
+    std::string context;
+    std::string grammar;
+};
+
+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, "  -pms N,     --prompt-ms N    [%-7d] prompt duration in milliseconds\n",             params.prompt_ms);
+    fprintf(stderr, "  -cms N,     --command-ms N   [%-7d] command duration in milliseconds\n",            params.command_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, "  -ng,        --no-gpu         [%-7s] disable GPU\n",                                 params.use_gpu ? "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, "  -cmd FNAME, --commands FNAME [%-7s] text file with allowed commands\n",             params.commands.c_str());
+    fprintf(stderr, "  -p,         --prompt         [%-7s] the required activation prompt\n",              params.prompt.c_str());
+    fprintf(stderr, "  -ctx,       --context        [%-7s] sample text to help the transcription\n",       params.context.c_str());
+    fprintf(stderr, "  --grammar-penalty N          [%-7.1f] scales down logits of nongrammar tokens\n",   params.grammar_penalty);
+    fprintf(stderr, "\n");
+}
+
+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 == "-pms" || arg == "--prompt-ms")     { params.prompt_ms     = std::stoi(argv[++i]); }
+        else if (arg == "-cms" || arg == "--command-ms")    { params.command_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 == "-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 == "-ng"  || arg == "--no-gpu")        { params.use_gpu       = false; }
+        else if (arg == "-l"   || arg == "--language")      { params.language      = argv[++i]; }
+        else if (arg == "-m"   || arg == "--model")         { params.model         = argv[++i]; }
+        else if (arg == "-f"   || arg == "--file")          { params.fname_out     = argv[++i]; }
+        else if (arg == "-cmd" || arg == "--commands")      { params.commands      = argv[++i]; }
+        else if (arg == "-p"   || arg == "--prompt")        { params.prompt        = argv[++i]; }
+        else if (arg == "-ctx" || arg == "--context")       { params.context       = argv[++i]; }
+        else if (                 arg == "--grammar-penalty") { params.grammar_penalty = std::stof(argv[++i]); }
+        else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            whisper_print_usage(argc, argv, params);
+            exit(0);
+        }
+    }
+
+    return true;
+}
+
+std::unique_ptr<WChess> g_wchess;
+int g_moveCount = 0;
+void set_move(const std::string & move, float) {
+    if (!move.empty()) {
+        g_moveCount++;
+        fprintf(stdout, "Move: %s\n\n", move.c_str());
+    }
+    else fprintf(stdout, "Move rejected\n\n");
+    fprintf(stdout, "%s\n", g_wchess->stringify_board().c_str());
+    fprintf(stdout, "%s\n", g_moveCount ? "White's turn" : "Black's turn");
+}
+
+audio_async g_audio(30*1000);
+bool g_listening = false;
+std::vector<float> g_pcmf32;
+
+bool read_input() {
+    std::string input;
+    while (true) {
+        fprintf(stdout, "[(l)isten/(p)ause/(q)uit]: ");
+        std::cin >> input;
+        fprintf(stdout, "\n");
+        if (input[0] == 'q') {
+            fprintf(stdout, "Quitting\n");
+            return false;
+        }
+        if (input[0] == 'l') {
+            if (!g_listening) {
+                fprintf(stdout, "Listening\n");
+                g_listening = true;
+                g_pcmf32.clear();
+                g_audio.resume();
+                g_audio.clear();
+            }
+            else fprintf(stdout, "Still listening\n");
+            return true;
+        }
+        else {
+            if (g_listening) {
+                g_listening = false;
+                g_audio.get(0, g_pcmf32);
+                g_audio.pause();
+                fprintf(stdout, "Processing\n");
+            }
+            else fprintf(stdout, "Not listening\n");
+            return true;
+        }
+    }
+    return true;
+}
+
+bool get_audio(std::vector<float> & pcmf32_cur) {
+    if (!read_input()) return false;
+    if (!g_pcmf32.empty()) pcmf32_cur = std::move(g_pcmf32);
+    else pcmf32_cur.clear();
+    return true;
+}
+
+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_params cparams;
+    cparams.use_gpu = params.use_gpu;
+
+    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
+    if (!ctx) {
+        fprintf(stderr, "%s: whisper_init_from_file_with_params() failed!\n", __func__);
+        return 1;
+    }
+
+    // init audio
+
+    if (!g_audio.init(params.capture_id, WHISPER_SAMPLE_RATE)) {
+        fprintf(stderr, "%s: audio.init() failed!\n", __func__);
+        return 1;
+    }
+
+    struct whisper_full_params wparams = whisper_full_default_params(whisper_sampling_strategy::WHISPER_SAMPLING_GREEDY);
+    wparams.offset_ms        = 0;
+    wparams.translate        = false;
+    wparams.no_context       = true;
+    wparams.single_segment   = true;
+    wparams.print_realtime   = false;
+    wparams.print_progress   = false;
+    wparams.print_timestamps = true;
+    wparams.print_special    = false;
+    wparams.no_timestamps    = true;
+
+    wparams.max_tokens       = 32;
+    wparams.audio_ctx        = 768; // partial encoder context for better performance
+
+    wparams.temperature     = 0.0f;
+    wparams.temperature_inc = 2.0f;
+    wparams.greedy.best_of  = 1;
+
+    wparams.beam_search.beam_size = 1;
+
+    wparams.language         = "en";
+
+    wparams.grammar_penalty = 100.0;
+
+    wparams.initial_prompt = params.context.data();
+
+    WChess::callbacks cb;
+    cb.get_audio = get_audio;
+    cb.set_move = set_move;
+
+    WChess::settings s;
+    s.vad_ms = 2000;
+    s.prompt_ms = params.prompt_ms;
+    s.command_ms = params.command_ms;
+    s.vad_thold = params.vad_thold;
+    s.freq_thold = params.freq_thold;
+    s.print_energy = params.print_energy;
+
+    g_wchess.reset(new WChess(ctx, wparams, cb, s));
+    set_move("start", 0);
+    g_wchess->run();
+
+    whisper_print_timings(ctx);
+    whisper_free(ctx);
+
+    return 0;
+}

examples/wchess/wchess.wasm/CMakeLists.txt

@@ -0,0 +1,51 @@
+set(TARGET wchess.wasm)
+
+add_executable(${TARGET}
+    wchess.wasm.cpp
+    )
+
+include(DefaultTargetOptions)
+
+target_link_libraries(${TARGET} PRIVATE
+    common
+    wchess-core
+    )
+
+unset(EXTRA_FLAGS)
+
+if (WHISPER_WASM_SINGLE_FILE)
+    set(EXTRA_FLAGS "-s SINGLE_FILE=1")
+    message(STATUS "Embedding WASM inside chess.js")
+
+    add_custom_command(
+        TARGET ${TARGET} POST_BUILD
+        COMMAND ${CMAKE_COMMAND} -E copy
+        ${CMAKE_BINARY_DIR}/bin/${TARGET}.js
+        ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${TARGET}/js/chess.js
+        )
+endif()
+
+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 FORCE_FILESYSTEM=1 \
+    -s EXPORTED_RUNTIME_METHODS=\"['print', 'printErr', 'ccall', 'cwrap']\" \
+    ${EXTRA_FLAGS} \
+    ")
+
+
+add_custom_command(
+        TARGET ${TARGET} POST_BUILD
+        COMMAND ${CMAKE_COMMAND} -E copy_directory
+        ${CMAKE_CURRENT_SOURCE_DIR}/chessboardjs-1.0.0
+        ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${TARGET}/
+        COMMAND ${CMAKE_COMMAND} -E copy
+        ${CMAKE_CURRENT_SOURCE_DIR}/jquery-3.7.1.min.js
+        ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${TARGET}/js/
+    )
+
+configure_file(${CMAKE_CURRENT_SOURCE_DIR}/index-tmpl.html  ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${TARGET}/index.html @ONLY)
+configure_file(${CMAKE_SOURCE_DIR}/examples/helpers.js    ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${TARGET}/js/helpers.js @ONLY)

examples/wchess/wchess.wasm/chessboardjs-1.0.0/css/chessboard-1.0.0.css

@@ -0,0 +1,54 @@
+/*! chessboard.js v1.0.0 | (c) 2019 Chris Oakman | MIT License chessboardjs.com/license */
+
+.clearfix-7da63 {
+  clear: both;
+}
+
+.board-b72b1 {
+  border: 2px solid #404040;
+  box-sizing: content-box;
+}
+
+.square-55d63 {
+  float: left;
+  position: relative;
+
+  /* disable any native browser highlighting */
+  -webkit-touch-callout: none;
+    -webkit-user-select: none;
+     -khtml-user-select: none;
+       -moz-user-select: none;
+        -ms-user-select: none;
+            user-select: none;
+}
+
+.white-1e1d7 {
+  background-color: #f0d9b5;
+  color: #b58863;
+}
+
+.black-3c85d {
+  background-color: #b58863;
+  color: #f0d9b5;
+}
+
+.highlight1-32417, .highlight2-9c5d2 {
+  box-shadow: inset 0 0 3px 3px yellow;
+}
+
+.notation-322f9 {
+  cursor: default;
+  font-family: "Helvetica Neue", Helvetica, Arial, sans-serif;
+  font-size: 14px;
+  position: absolute;
+}
+
+.alpha-d2270 {
+  bottom: 1px;
+  right: 3px;
+}
+
+.numeric-fc462 {
+  top: 2px;
+  left: 2px;
+}

examples/wchess/wchess.wasm/chessboardjs-1.0.0/css/chessboard-1.0.0.min.css

@@ -0,0 +1,2 @@
+/*! chessboard.js v1.0.0 | (c) 2019 Chris Oakman | MIT License chessboardjs.com/license */
+.clearfix-7da63{clear:both}.board-b72b1{border:2px solid #404040;box-sizing:content-box}.square-55d63{float:left;position:relative;-webkit-touch-callout:none;-webkit-user-select:none;-khtml-user-select:none;-moz-user-select:none;-ms-user-select:none;user-select:none}.white-1e1d7{background-color:#f0d9b5;color:#b58863}.black-3c85d{background-color:#b58863;color:#f0d9b5}.highlight1-32417,.highlight2-9c5d2{box-shadow:inset 0 0 3px 3px #ff0}.notation-322f9{cursor:default;font-family:"Helvetica Neue",Helvetica,Arial,sans-serif;font-size:14px;position:absolute}.alpha-d2270{bottom:1px;right:3px}.numeric-fc462{top:2px;left:2px}

examples/wchess/wchess.wasm/chessboardjs-1.0.0/js/chessboard-1.0.0/CHANGELOG.md

@@ -0,0 +1,32 @@
+# chessboard.js Change Log
+
+All notable changes to this project will be documented in this file.
+
+## [1.0.0] - 2019-06-11
+- Orientation methods now return current orientation. [Issue #64]
+- Drop support for IE8
+- Do not check for `window.JSON` (Error #1004)
+- Rename `ChessBoard` to `Chessboard` (`ChessBoard` is still supported, however)
+- id query selectors are now supported as the first argument to `Chessboard()`
+- Remove Error #1002
+- Format code according to [StandardJS]
+- Bump minimum jQuery version to 1.8.3
+- Throttle piece drag functions
+
+## [0.3.0] - 2013-08-10
+- Added `appearSpeed` animation config property
+- Added `onSnapbackEnd` event
+- Added `onMoveEnd` event
+
+## [0.2.0] - 2013-08-05
+- Added `onMouseoverSquare` and `onMouseoutSquare` events
+- Added `onSnapEnd` event
+- Added square code as CSS class on the squares
+- Added [chess.js] integration examples
+
+## [0.1.0] - 2013-05-21
+- Initial release
+
+[chess.js]:https://github.com/jhlywa/chess.js
+[Issue #64]:https://github.com/oakmac/chessboardjs/issues/64
+[StandardJS]:https://standardjs.com/

examples/wchess/wchess.wasm/chessboardjs-1.0.0/js/chessboard-1.0.0/LICENSE.md

@@ -0,0 +1,20 @@
+Copyright 2019 Chris Oakman
+
+Permission is hereby granted, free of charge, to any person obtaining
+a copy of this software and associated documentation files (the
+"Software"), to deal in the Software without restriction, including
+without limitation the rights to use, copy, modify, merge, publish,
+distribute, sublicense, and/or sell copies of the Software, and to
+permit persons to whom the Software is furnished to do so, subject to
+the following conditions:
+
+The above copyright notice and this permission notice shall be
+included in all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE
+LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
+OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
+WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

examples/wchess/wchess.wasm/chessboardjs-1.0.0/js/chessboard-1.0.0/README.md

@@ -0,0 +1,82 @@
+# chessboard.js
+
+chessboard.js is a JavaScript chessboard component. It depends on [jQuery].
+
+Please see [chessboardjs.com] for documentation and examples.
+
+## What is chessboard.js?
+
+chessboard.js is a JavaScript chessboard component with a flexible "just a
+board" API that
+
+chessboard.js is a standalone JavaScript Chess Board. It is designed to be "just
+a board" and expose a powerful API so that it can be used in different ways.
+Here's a non-exhaustive list of things you can do with chessboard.js:
+
+- Use chessboard.js to show game positions alongside your expert commentary.
+- Use chessboard.js to have a tactics website where users have to guess the best
+  move.
+- Integrate chessboard.js and [chess.js] with a PGN database and allow people to
+  search and playback games (see [Example 5000])
+- Build a chess server and have users play their games out using the
+  chessboard.js board.
+
+chessboard.js is flexible enough to handle any of these situations with relative
+ease.
+
+## What can chessboard.js **not** do?
+
+The scope of chessboard.js is limited to "just a board." This is intentional and
+makes chessboard.js flexible for handling a multitude of chess-related problems.
+
+This is a common source of confusion for new users. [remove?]
+
+Specifically, chessboard.js does not understand anything about how the game of
+chess is played: how a knight moves, who's turn is it, is White in check?, etc.
+
+Fortunately, the powerful [chess.js] library deals with exactly this sort of
+problem domain and plays nicely with chessboard.js's flexible API. Some examples
+of chessboard.js combined with chess.js: 5000, 5001, 5002
+
+Please see the powerful [chess.js] library for an API to deal with these sorts
+of questions.
+
+
+This logic is distinct from the logic of the board. Please see the powerful
+[chess.js] library for this aspect of your application.
+
+
+
+Here is a list of things that chessboard.js is **not**:
+
+- A chess engine
+- A legal move validator
+- A PGN parser
+
+chessboard.js is designed to work well with any of those things, but the idea
+behind chessboard.js is that the logic that controls the board should be
+independent of those other problems.
+
+## Docs and Examples
+
+- Docs - <http://chessboardjs.com/docs>
+- Examples - <http://chessboardjs.com/examples>
+
+## Developer Tools
+
+```sh
+# create a build in the build/ directory
+npm run build
+
+# re-build the website
+npm run website
+```
+
+## License
+
+[MIT License](LICENSE.md)
+
+[jQuery]:https://jquery.com/
+[chessboardjs.com]:http://chessboardjs.com
+[chess.js]:https://github.com/jhlywa/chess.js
+[Example 5000]:http://chessboardjs.com/examples#5000

examples/wchess/wchess.wasm/chessboardjs-1.0.0/js/chessboard-1.0.0/package.json

@@ -0,0 +1,29 @@
+{
+  "author": "Chris Oakman <chris@oakmac.com> (http://chrisoakman.com/)",
+  "name": "@chrisoakman/chessboardjs",
+  "description": "JavaScript chessboard widget",
+  "homepage": "https://chessboardjs.com",
+  "license": "MIT",
+  "version": "1.0.0",
+  "repository": {
+    "type": "git",
+    "url": "git://github.com/oakmac/chessboardjs.git"
+  },
+  "files": ["dist/"],
+  "dependencies": {
+    "jquery": ">=3.4.1"
+  },
+  "devDependencies": {
+    "csso": "3.5.1",
+    "fs-plus": "3.1.1",
+    "kidif": "1.1.0",
+    "mustache": "2.3.0",
+    "standard": "10.0.2",
+    "uglify-js": "3.6.0"
+  },
+  "scripts": {
+    "build": "standard lib/chessboard.js && node scripts/build.js",
+    "standard": "standard --fix lib/*.js website/js/*.js",
+    "website": "node scripts/website.js"
+  }
+}

examples/wchess/wchess.wasm/index-tmpl.html

@@ -0,0 +1,499 @@
+<!doctype html>
+<html lang="en-us">
+    <head>
+        <title>wchess : voice-controlled chess using Whisper + WebAssembly</title>
+        <script src="https://cdnjs.cloudflare.com/ajax/libs/iframe-resizer/4.3.1/iframeResizer.contentWindow.min.js"></script>
+
+        <meta name="viewport" content="width=device-width, initial-scale=0.7, maximum-scale=1, minimum-scale=0.7, user-scalable=no"/>
+        <meta name="apple-mobile-web-app-capable" content="yes" />
+
+        <style>
+            #output {
+                width: 100%;
+                height: 100%;
+                margin: 0 auto;
+                margin-top: 10px;
+                border-left: 0px;
+                border-right: 0px;
+                padding-left: 0px;
+                padding-right: 0px;
+                display: block;
+                background-color: black;
+                color: white;
+                font-size: 10px;
+                font-family: 'Lucida Console', Monaco, monospace;
+                outline: none;
+                white-space: pre;
+                overflow-wrap: normal;
+                overflow-x: scroll;
+            }
+            .button {
+                background-color: #000000;
+                color: #FFFFFF;
+                padding: 20px;
+                border-radius: 10px;
+                -moz-border-radius: 10px;
+                -webkit-border-radius: 10px;
+                margin:10px;
+                width:  100px;
+                height:  50px;
+                -webkit-touch-callout: none; /* Safari */
+                -webkit-user-select: none; /* Chrome */
+                -moz-user-select: none; /* Firefox */
+                -ms-user-select: none; /* Internet Explorer/Edge */
+                user-select: none;
+            }
+            button[disabled]{
+                background-color: #cccccc;
+                color: #666666;
+                padding: 20px;
+                border-radius: 10px;
+                -moz-border-radius: 10px;
+                -webkit-border-radius: 10px;
+                margin:10px;
+                width: 100px;
+            }
+            .center {
+                display: flex;
+                justify-content: center;
+                align-items: center;
+                width: 500px;
+            }
+            #description {
+                width: 500px;
+            }
+        </style>
+        <link rel="stylesheet" href="css/chessboard-1.0.0.min.css" integrity="sha384-q94+BZtLrkL1/ohfjR8c6L+A6qzNH9R2hBLwyoAfu3i/WCvQjzL2RQJ3uNHDISdU" crossorigin="anonymous">
+    </head>
+    <body>
+        <div id="main-container">
+            <div id="description">
+                <b>wchess : voice-controlled chess using Whisper + WebAssembly</b>
+
+                <br><br>
+
+                This is a demonstration of using Whisper to recognize voice commands in the browser.
+
+                <br><br>
+
+                Usage:<br>
+
+                <ul>
+                    <li>Select a Whisper model</li>
+                    <li>Accept the microphone permission request if prompted</li>
+                    <li>Hold the button and say a chess move (e.g. "Knight to c3")</li>
+                    <li>Release the button and wait for the move to be recognized</li>
+                    <li>Repeat</li>
+                </ul>
+
+                Examples:<br>
+
+                <ul>
+                    <li><b>"d4"</b></li>
+                    <li><b>"e2 e4"</b></li>
+                    <li><b>"Knight f3"</b></li>
+                    <li><b>"Bishop to b5"</b></li>
+                </ul>
+
+                Features:<br>
+
+                <ul>
+                    <li>Model quantization for reduced memory footprint (~42MB)</li>
+                    <li><a href="https://github.com/ggerganov/whisper.cpp/pull/1229">Grammar-based sampling</a> for improved recognition accuracy</li>
+                </ul>
+
+                <b>
+                Note that not all chess moves are supported. For example, castling and pawn promotion
+                currently do not work, but can be easily implemented. There could also be some bugs in
+                the move handling logic in general. The main reason for that is to keep the implementation
+                simple. The assumption is that a real application would already have a proper move
+                validation logic in place.<br><br>
+
+                The main purpose of this example is to demonstrate the capabilities of whisper.cpp and
+                its application in the browser for voice recognition locally on your device.
+                </b>
+
+                <br><br>
+
+                You can find more about this project on <a href="https://github.com/ggerganov/whisper.cpp/tree/master/examples/wchess">GitHub</a>.
+
+                <br><br>
+
+                <b>More examples:</b>
+                    <a href="https://whisper.ggerganov.com/">main</a> |
+                    <a href="https://whisper.ggerganov.com/bench">bench</a> |
+                    <a href="https://whisper.ggerganov.com/stream">stream</a> |
+                    <a href="https://whisper.ggerganov.com/command">command</a> |
+                    <a href="https://whisper.ggerganov.com/talk">talk</a> |
+
+                <br><br>
+
+            </div>
+
+            <hr>
+
+            <div id="model-whisper">
+                Whisper model: <span id="model-whisper-status"></span>
+                <button id="fetch-whisper-tiny-en" onclick="loadWhisper()">tiny.en (Q8_0, 42 MB)</button>
+                <span id="fetch-whisper-progress"></span>
+                <br><br>
+                <button id="clear" onclick="clearCache()">Clear browser cache</button>
+                <!--
+                    <input type="file" id="file" name="file" onchange="loadFile(event, 'whisper.bin')" />
+                -->
+            </div>
+
+            <div id="game">
+                <br>
+                <div id="chessboard" style="width: 500px"></div>
+                <script src="js/jquery-3.7.1.min.js"></script>
+                <script src="js/chessboard-1.0.0.min.js"></script>
+                <script>
+                    var board = Chessboard('chessboard', 'start')
+                    var move_count = 0;
+                </script>
+
+                <br>
+
+                <div id="state">
+                    Status: <b><span id="state-status">select model</span></b>
+
+                    <div id="input" class="center">
+                        <button id="toggler" class="button" onselectstart="return false" style="display: none">Hold</button>
+                    </div>
+
+                    <pre id="state-grammar">[The grammar will be displayed here]</pre>
+
+                    <pre id="state-moves">[The moves will be displayed here]</pre>
+                </div>
+            </div>
+
+            <hr>
+
+            Debug output:
+            <textarea id="output" rows="20"></textarea>
+
+            <br>
+
+            <b>Troubleshooting</b>
+
+            <br><br>
+
+            The page does some heavy computations, so make sure:
+
+            <ul>
+                <li>To use a modern web browser (e.g. Chrome, Firefox)</li>
+                <li>Your browser supports WASM <a href="https://webassembly.org/roadmap/">Fixed-width SIMD</a></li>
+            </ul>
+
+            <div class="cell-version">
+                <span>
+                    |
+                    Build time: <span class="nav-link">@GIT_DATE@</span> |
+                    Commit hash: <a class="nav-link" href="https://github.com/ggerganov/whisper.cpp/commit/@GIT_SHA1@">@GIT_SHA1@</a> |
+                    Commit subject: <span class="nav-link">@GIT_COMMIT_SUBJECT@</span> |
+                    <a class="nav-link" href="https://github.com/ggerganov/whisper.cpp/tree/master/examples/command.wasm">Source Code</a> |
+                </span>
+            </div>
+        </div>
+
+        <script type="text/javascript" src="js/helpers.js"></script>
+        <script type='text/javascript'>
+            // web audio context
+            var context = null;
+
+            // the command instance
+            var instance = null;
+
+            // model name
+            var model_whisper = null;
+            var model_file = null;
+
+            var module_ready = null;
+
+            var Module = {
+                print: printTextarea,
+                printErr: printTextarea,
+                setStatus: function(text) {
+                    printTextarea('js: ' + text);
+                },
+                monitorRunDependencies: function(left) {
+                },
+                preRun: function() {
+                    printTextarea('js: Preparing ...');
+                },
+                postRun: function() {
+                    printTextarea('js: Module initialized successfully!');
+                    module_ready = true;
+                    initInstance();
+                }
+            };
+
+            function initInstance() {
+                if (!module_ready || !model_file || instance) return
+
+                instance = Module.init(model_file);
+
+                if (instance) {
+                    setStatus('Ready');
+                    printTextarea("js: whisper initialized, instance: " + instance);
+                }
+                else {
+                    printTextarea("js: failed to initialize whisper");
+                }
+            }
+
+            function setStatus(text) {
+                document.getElementById('state-status').innerHTML = text;
+            }
+
+            //
+            // fetch models
+            //
+
+            let dbVersion = 1
+            let dbName    = 'whisper.ggerganov.com';
+            let indexedDB = window.indexedDB || window.mozIndexedDB || window.webkitIndexedDB || window.msIndexedDB
+
+            function storeFS(fname, buf) {
+                // write to WASM file using FS_createDataFile
+                // if the file exists, delete it
+                try {
+                    Module.FS_unlink(fname);
+                } catch (e) {
+                    // ignore
+                }
+
+                Module.FS_createDataFile("/", fname, buf, true, true);
+
+                printTextarea('storeFS: stored model: ' + fname + ' size: ' + buf.length);
+
+                document.getElementById('model-whisper-status').innerHTML = 'loaded "' + model_whisper + '"!';
+
+                model_file = fname;
+                initInstance();
+            }
+
+            function loadWhisper() {
+                setStatus('Loading')
+                //let url     = 'https://whisper.ggerganov.com/ggml-model-whisper-tiny.en-q8_0.bin';
+                let url     = 'https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-tiny.en-q8_0.bin';
+                let dst     = 'whisper.bin';
+                let size_mb = 42;
+
+                model_whisper = 'tiny.en-q8_0';
+
+                document.getElementById('model-whisper-status').innerHTML = 'loading "' + model_whisper + '" ... ';
+                document.getElementById('fetch-whisper-tiny-en').style.display = 'none';
+
+                cbProgress = function(p) {
+                    let el = document.getElementById('fetch-whisper-progress');
+                    el.innerHTML = Math.round(100*p) + '%';
+                };
+
+                cbCancel = function() {
+                    var el;
+                    el = document.getElementById('model-whisper-status');  if (el) el.innerHTML = '';
+                };
+
+                loadRemote(url, dst, size_mb, cbProgress, storeFS, cbCancel, printTextarea);
+
+                // init audio capture so that the user receives a permission request
+                {
+                    let context = new AudioContext({
+                        sampleRate: 16000,
+                        channelCount: 1,
+                        echoCancellation: false,
+                        autoGainControl:  true,
+                        noiseSuppression: true,
+                    });
+                    navigator.mediaDevices.getUserMedia({audio: true, video: false})
+                        .then(function(s) {
+                            stream = s;
+                            stream.getTracks().forEach(function(track) {
+                                track.stop();
+                            });
+                        })
+                        .catch(function(err) {
+                            printTextarea('js: error getting audio stream: ' + err);
+                        });
+                    context.close();
+                }
+
+                document.getElementById('toggler').style.display = 'block';
+            }
+
+            //
+            // microphone
+            //
+
+            const kSampleRate = 16000;
+            const kRestartRecording_s = 120;
+            const kIntervalAudio_ms = 250; // pass the recorded audio to the C++ instance at this rate
+
+            var mediaRecorder = null;
+            var doRecording = false;
+            var startTime = 0;
+
+            window.AudioContext = window.AudioContext || window.webkitAudioContext;
+            window.OfflineAudioContext = window.OfflineAudioContext || window.webkitOfflineAudioContext;
+
+            function stopRecording() {
+                if (mediaRecorder) {
+                    mediaRecorder.stop();
+                }
+            }
+
+            function startRecording() {
+                if (!context) {
+                    context = new AudioContext({
+                        sampleRate: kSampleRate,
+                        channelCount: 1,
+                        echoCancellation: false,
+                        autoGainControl:  true,
+                        noiseSuppression: true,
+                    });
+                }
+
+                startTime = Date.now();
+
+                var chunks = [];
+                var stream = null;
+
+                navigator.mediaDevices.getUserMedia({audio: true, video: false})
+                    .then(function(s) {
+                        stream = s;
+                        mediaRecorder = new MediaRecorder(stream);
+                        mediaRecorder.ondataavailable = function(e) {
+                            chunks.push(e.data);
+
+                            var blob = new Blob(chunks, { 'type' : 'audio/ogg; codecs=opus' });
+                            var reader = new FileReader();
+
+                            reader.onload = function(event) {
+                                var buf = new Uint8Array(reader.result);
+                                context.decodeAudioData(buf.buffer, function(audioBuffer) {
+                                    var offlineContext = new OfflineAudioContext(audioBuffer.numberOfChannels, audioBuffer.length, audioBuffer.sampleRate);
+                                    var source = offlineContext.createBufferSource();
+                                    source.buffer = audioBuffer;
+                                    source.connect(offlineContext.destination);
+                                    source.start(0);
+
+                                    offlineContext.startRendering().then(function(renderedBuffer) {
+                                        let audio = renderedBuffer.getChannelData(0);
+                                        printTextarea('js: number of samples: ' + audio.length);
+                                        Module.set_audio(instance, audio);
+                                    });
+
+                                    mediaRecorder = null;
+                                    context = null;
+                                });
+                            }
+
+                            reader.readAsArrayBuffer(blob);
+                        };
+
+                        mediaRecorder.onstop = function(e) {
+                            stream.getTracks().forEach(function(track) {
+                                track.stop();
+                            });
+                        };
+
+                        mediaRecorder.start();
+                    })
+                    .catch(function(err) {
+                        printTextarea('js: error getting audio stream: ' + err);
+                    });
+            }
+
+            //
+            // main
+            //
+
+            var nLines = 0;
+            var movesAll = '';
+
+            // document.body.addEventListener('keydown', function(event) {
+            //     if (event.keyCode === 32) {
+            //         document.getElementById('toggler').innerText = "";
+            //         onStart();
+            //     }
+            // }, true);
+
+            // document.body.addEventListener('keyup', function(event) {
+            //     if (event.keyCode === 32) {
+            //         document.getElementById('toggler').innerText = "Hold";
+            //         onStop();
+            //     }
+            // }, true);
+
+            document.getElementById('toggler').addEventListener("touchstart", function(event){
+                this.innerText = "";
+                onStart();
+            }, true);
+
+            document.getElementById('toggler').addEventListener("touchend", function(event){
+                this.innerText = "Hold";
+                onStop();
+            }, true)
+
+            document.getElementById('toggler').addEventListener('mousedown', function(event) {
+                this.innerText = "";
+                onStart();
+            }, true);
+
+            document.getElementById('toggler').addEventListener('mouseup', function(event) {
+                this.innerText = "Hold";
+                onStop();
+            }, true);
+
+            function onStart() {
+                if (!instance) return;
+                setStatus('Listening');
+
+                startRecording();
+            }
+
+            function onStop() {
+                setStatus('Processing');
+                printTextarea('js: stopping recording ...');
+                stopRecording();
+            }
+
+            function setMove(move, prob) {
+                if (move != null && move.length > 1) {
+                    let gameOver =  move[move.length - 1] === '#';
+                    if (gameOver) {
+                        move = move.substring(0, move.length - 1);
+                        document.getElementById('toggler').disabled = true;
+                    }
+                    board.move(move);
+
+                    movesAll += move + ', prob = ' + prob.toFixed(2) + '% <br>';
+                    nLines++;
+
+                    // if more than 10 lines, remove the first line
+                    if (nLines > 10) {
+                        var i = movesAll.indexOf('<br>');
+                        if (i > 0) {
+                            movesAll = movesAll.substring(i + 4);
+                            nLines--;
+                        }
+                    }
+                    ++move_count;
+                    setStatus(gameOver ? 'Done' : move_count % 2 ? 'Black\'s turn' : 'White\'s turn');
+                    document.getElementById('state-moves').innerHTML = movesAll;
+                }
+                else {
+                    setStatus('Failed. ' + (move_count % 2 ? 'Black\'s turn' : 'White\'s turn'));
+                }
+            }
+
+            function setGrammar(grammar) {
+                document.getElementById('state-grammar').innerHTML = grammar;
+            }
+
+        </script>
+        <script type="text/javascript" src="js/chess.js"></script>
+    </body>
+</html>

examples/wchess/wchess.wasm/wchess.wasm.cpp

@@ -0,0 +1,141 @@
+#include <WChess.h>
+#include <emscripten.h>
+#include <emscripten/bind.h>
+
+#include <thread>
+
+constexpr int N_THREAD = 8;
+
+std::vector<struct whisper_context *> g_contexts(4, nullptr);
+
+std::mutex  g_mutex;
+std::thread g_worker;
+
+std::condition_variable g_cv;
+
+bool g_running(false);
+std::vector<float> g_pcmf32;
+
+void set_move(const std::string & move, float prob) {
+    MAIN_THREAD_EM_ASM({
+        setMove(UTF8ToString($0), $1)
+    }, move.c_str(), prob);
+}
+
+void set_grammar(const std::string & grammar) {
+    MAIN_THREAD_EM_ASM({
+        setGrammar(UTF8ToString($0))
+    }, grammar.c_str());
+}
+
+bool get_audio(std::vector<float> & audio) {
+    std::unique_lock<std::mutex> lock(g_mutex);
+    g_cv.wait(lock, [] { return !g_running || !g_pcmf32.empty(); });
+    if (!g_running) return false;
+    audio = std::move(g_pcmf32);
+    return true;
+}
+
+void wchess_main(size_t i) {
+    struct whisper_full_params wparams = whisper_full_default_params(whisper_sampling_strategy::WHISPER_SAMPLING_GREEDY);
+
+    wparams.n_threads        = std::min(N_THREAD, (int) std::thread::hardware_concurrency());
+    wparams.offset_ms        = 0;
+    wparams.translate        = false;
+    wparams.no_context       = true;
+    wparams.single_segment   = true;
+    wparams.print_realtime   = false;
+    wparams.print_progress   = false;
+    wparams.print_timestamps = true;
+    wparams.print_special    = false;
+    wparams.no_timestamps    = true;
+
+    wparams.max_tokens       = 32;
+    wparams.audio_ctx        = 1280; // partial encoder context for better performance
+
+    wparams.temperature      = 0.0f;
+    wparams.temperature_inc  = 2.0f;
+    wparams.greedy.best_of   = 1;
+
+    wparams.beam_search.beam_size = 1;
+
+    wparams.language         = "en";
+
+    wparams.grammar_penalty = 100.0;
+    wparams.initial_prompt = "bishop to c3, rook to d4, knight to e5, d4 d5, knight to c3, c3, queen to d4, king b1, pawn to a1, bishop to b2, knight to c3,";
+
+    printf("command: using %d threads\n", wparams.n_threads);
+
+    WChess::callbacks cb;
+    cb.get_audio = get_audio;
+    cb.set_move = set_move;
+    cb.set_grammar = set_grammar;
+
+    WChess(g_contexts[i], wparams, cb, {}).run();
+
+    if (i < g_contexts.size()) {
+        whisper_free(g_contexts[i]);
+        g_contexts[i] = nullptr;
+    }
+}
+
+EMSCRIPTEN_BINDINGS(command) {
+    emscripten::function("init", emscripten::optional_override([](const std::string & path_model) {
+        for (size_t i = 0; i < g_contexts.size(); ++i) {
+            if (g_contexts[i] == nullptr) {
+                g_contexts[i] = whisper_init_from_file_with_params(path_model.c_str(), whisper_context_default_params());
+                if (g_contexts[i] != nullptr) {
+                    g_running = true;
+                    if (g_worker.joinable()) {
+                        g_worker.join();
+                    }
+                    g_worker = std::thread([i]() {
+                        wchess_main(i);
+                    });
+
+                    return i + 1;
+                } else {
+                    return (size_t) 0;
+                }
+            }
+        }
+
+        return (size_t) 0;
+    }));
+
+    emscripten::function("free", emscripten::optional_override([](size_t /* index */) {
+        {
+            std::unique_lock<std::mutex> lock(g_mutex);
+            g_running = false;
+        }
+        g_cv.notify_one();
+    }));
+
+    emscripten::function("set_audio", emscripten::optional_override([](size_t index, const emscripten::val & audio) {
+        --index;
+
+        if (index >= g_contexts.size()) {
+            return -1;
+        }
+
+        if (g_contexts[index] == nullptr) {
+            return -2;
+        }
+
+        {
+            std::lock_guard<std::mutex> lock(g_mutex);
+            const int n = audio["length"].as<int>();
+
+            emscripten::val heap = emscripten::val::module_property("HEAPU8");
+            emscripten::val memory = heap["buffer"];
+
+            g_pcmf32.resize(n);
+
+            emscripten::val memoryView = audio["constructor"].new_(memory, reinterpret_cast<uintptr_t>(g_pcmf32.data()), n);
+            memoryView.call<void>("set", audio);
+        }
+        g_cv.notify_one();
+
+        return 0;
+    }));
+}

examples/whisper.objc/whisper.objc/ViewController.m

@@ -206,6 +206,7 @@ void AudioInputCallback(void * inUserData,
         params.offset_ms        = 0;
         params.no_context       = true;
         params.single_segment   = self->stateInp.isRealtime;
+        params.no_timestamps    = params.single_segment;
 
         CFTimeInterval startTime = CACurrentMediaTime();
 

ggml-alloc.c

@@ -137,7 +137,7 @@ void ggml_tallocr_alloc(ggml_tallocr_t alloc, struct ggml_tensor * tensor) {
 
 #ifdef GGML_ALLOCATOR_DEBUG
     add_allocated_tensor(alloc, tensor);
-    size_t cur_max = (char*)addr - (char*)alloc->data + size;
+    size_t cur_max = (char*)addr - (char*)alloc->base + size;
     if (cur_max > alloc->max_size) {
         printf("max_size = %.2f MB: tensors: ", cur_max / 1024.0 / 1024.0);
         for (int i = 0; i < 1024; i++) {
@@ -168,10 +168,6 @@ static void ggml_tallocr_free_tensor(ggml_tallocr_t alloc, struct ggml_tensor *
     size = aligned_offset(NULL, size, alloc->alignment);
     AT_PRINTF("%s: freeing %s at %p (%zu bytes) - n_free_blocks = %d\n", __func__, tensor->name, ptr, size, alloc->n_free_blocks);
 
-    if (!alloc->measure) {
-        ggml_backend_buffer_free_tensor(alloc->buffer, tensor);
-    }
-
 #ifdef GGML_ALLOCATOR_DEBUG
     remove_allocated_tensor(alloc, tensor);
 #endif
@@ -237,7 +233,7 @@ void ggml_tallocr_reset(ggml_tallocr_t alloc) {
 }
 
 ggml_tallocr_t ggml_tallocr_new(void * data, size_t size, size_t alignment) {
-    struct ggml_backend_buffer * buffer = ggml_backend_cpu_buffer_from_ptr(NULL, data, size);
+    struct ggml_backend_buffer * buffer = ggml_backend_cpu_buffer_from_ptr(data, size);
 
     ggml_tallocr_t alloc = (ggml_tallocr_t)malloc(sizeof(struct ggml_tallocr));
 
@@ -449,7 +445,6 @@ static ggml_tallocr_t node_tallocr(ggml_gallocr_t galloc, struct ggml_tensor * n
 static void init_view(ggml_gallocr_t galloc, struct ggml_tensor * view, bool update_backend) {
     ggml_tallocr_t alloc = node_tallocr(galloc, view);
 
-    //printf("init_view: %s from src %s\n", view->name, view->view_src->name);
     GGML_ASSERT(view->view_src != NULL && view->view_src->data != NULL);
     if (update_backend) {
         view->backend = view->view_src->backend;
@@ -459,7 +454,7 @@ static void init_view(ggml_gallocr_t galloc, struct ggml_tensor * view, bool upd
 
     // FIXME: the view should be initialized by the owning buffer, but currently this breaks the CUDA backend
     // due to the ggml_tensor_extra_gpu ring buffer overwriting the KV cache extras
-    assert(ggml_tallocr_is_measure(alloc) || !view->buffer || view->buffer->backend == alloc->buffer->backend);
+    assert(ggml_tallocr_is_measure(alloc) || !view->buffer || view->buffer->buft == alloc->buffer->buft);
 
     if (!alloc->measure) {
         ggml_backend_buffer_init_tensor(alloc->buffer, view);
@@ -765,3 +760,43 @@ size_t ggml_allocr_max_size(ggml_allocr_t alloc) {
 size_t ggml_allocr_alloc_graph(ggml_allocr_t alloc, struct ggml_cgraph * graph) {
     return ggml_gallocr_alloc_graph(alloc->galloc, alloc->talloc, graph);
 }
+
+// utils
+ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft) {
+    GGML_ASSERT(ggml_get_no_alloc(ctx) == true);
+
+    size_t alignment = ggml_backend_buft_get_alignment(buft);
+
+    size_t nbytes = 0;
+    for (struct ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+        if (t->data == NULL && t->view_src == NULL) {
+            nbytes += GGML_PAD(ggml_backend_buft_get_alloc_size(buft, t), alignment);
+        }
+    }
+
+    if (nbytes == 0) {
+        fprintf(stderr, "%s: no tensors to allocate\n", __func__);
+        return NULL;
+    }
+
+    ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(buft, nbytes);
+    ggml_tallocr_t tallocr = ggml_tallocr_new_from_buffer(buffer);
+
+    for (struct ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+        if (t->data == NULL) {
+            if (t->view_src == NULL) {
+                ggml_tallocr_alloc(tallocr, t);
+            } else {
+                ggml_backend_view_init(buffer, t);
+            }
+        }
+    }
+
+    ggml_tallocr_free(tallocr);
+
+    return buffer;
+}
+
+ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors(struct ggml_context * ctx, ggml_backend_t backend) {
+    return ggml_backend_alloc_ctx_tensors_from_buft(ctx, ggml_backend_get_default_buffer_type(backend));
+}

ggml-alloc.h

@@ -8,6 +8,7 @@ extern "C" {
 
 struct ggml_backend;
 struct ggml_backend_buffer;
+struct ggml_backend_buffer_type;
 
 //
 // Legacy API
@@ -42,7 +43,7 @@ GGML_API size_t ggml_allocr_alloc_graph(ggml_allocr_t alloc, struct ggml_cgraph
 // ggml-backend v2 API
 //
 
-// Seperate tensor and graph allocator objects
+// Separate tensor and graph allocator objects
 // This is necessary for multi-backend allocation because the graph allocator needs to use multiple tensor allocators
 // The original API is kept as a wrapper around the new API
 
@@ -80,6 +81,12 @@ GGML_API void   ggml_gallocr_alloc_graph_n(
                     struct ggml_hash_set hash_set,
                     ggml_tallocr_t * hash_node_talloc);
 
+
+// Utils
+// Create a buffer and allocate all the tensors in a ggml_context
+GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, struct ggml_backend_buffer_type * buft);
+GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors(struct ggml_context * ctx, struct ggml_backend * backend);
+
 #ifdef  __cplusplus
 }
 #endif

ggml-backend-impl.h

@@ -12,31 +12,50 @@ extern "C" {
     // Backend buffer
     //
 
+    // buffer type
+    typedef void * ggml_backend_buffer_type_context_t;
+
+    struct ggml_backend_buffer_type_i {
+        ggml_backend_buffer_t (*alloc_buffer)    (ggml_backend_buffer_type_t buft, size_t size);
+        size_t                (*get_alignment)   (ggml_backend_buffer_type_t buft); // tensor alignment
+        size_t                (*get_alloc_size)  (ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor); // data size needed to allocate the tensor, including padding
+        bool                  (*supports_backend)(ggml_backend_buffer_type_t buft, ggml_backend_t backend); // check if the buffer type is usable by the backend
+    };
+
+    struct ggml_backend_buffer_type {
+        struct ggml_backend_buffer_type_i  iface;
+        ggml_backend_buffer_type_context_t context;
+    };
+
+    // buffer
     typedef void * ggml_backend_buffer_context_t;
 
     struct ggml_backend_buffer_i {
-        void   (*free_buffer)   (ggml_backend_buffer_t buffer);
+        void     (*free_buffer)(ggml_backend_buffer_t buffer);
-        void * (*get_base)      (ggml_backend_buffer_t buffer); // get base pointer
+        //void     (*reset)      (ggml_backend_buffer_t buffer); // reset any internal state due to tensor initialization, such as tensor extras
-        size_t (*get_alloc_size)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // pre-allocation callback
+        void *   (*get_base)   (ggml_backend_buffer_t buffer);
-        void   (*init_tensor)   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // post-allocation callback
+        void     (*init_tensor)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-        void   (*free_tensor)   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // pre-free callback
+        void     (*set_tensor) (ggml_backend_buffer_t buffer,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+        void     (*get_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+        // (optional) copy tensor between different buffer-type, allow for single-copy tranfers
+        void (*cpy_tensor_from)(ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst);
+        void (*cpy_tensor_to)  (ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst);
     };
 
     struct ggml_backend_buffer {
         struct ggml_backend_buffer_i  iface;
-
+        ggml_backend_buffer_type_t    buft;
-        ggml_backend_t                backend;
         ggml_backend_buffer_context_t context;
-
         size_t size;
     };
 
-    GGML_API ggml_backend_buffer_t ggml_backend_buffer_init(
+    ggml_backend_buffer_t ggml_backend_buffer_init(
-            struct ggml_backend                  * backend,
+                   ggml_backend_buffer_type_t      buft,
             struct ggml_backend_buffer_i           iface,
                    ggml_backend_buffer_context_t   context,
                    size_t                          size);
 
+
     //
     // Backend
     //
@@ -49,20 +68,17 @@ extern "C" {
         void (*free)(ggml_backend_t backend);
 
         // buffer allocation
-        ggml_backend_buffer_t (*alloc_buffer)(ggml_backend_t backend, size_t size);
+        ggml_backend_buffer_type_t (*get_default_buffer_type)(ggml_backend_t backend);
 
-        // get buffer alignment
+        // (optional) asynchroneous tensor data access
-        size_t (*get_alignment)(ggml_backend_t backend);
-
-        // tensor data access
-        // these functions can be asynchronous, helper functions are provided for synchronous access that automatically call synchronize
         void (*set_tensor_async)(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
         void (*get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-        void (*synchronize)     (ggml_backend_t backend);
 
-        // (optional) copy tensor between different backends, allow for single-copy tranfers
+        // (optional) asynchroneous tensor copy
-        void (*cpy_tensor_from)(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
+        void (*cpy_tensor_from_async)(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
-        void (*cpy_tensor_to)  (ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
+        void (*cpy_tensor_to_async)  (ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
+
+        void (*synchronize)     (ggml_backend_t backend);
 
         // compute graph with a plan
         ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, struct ggml_cgraph * cgraph);
@@ -82,6 +98,15 @@ extern "C" {
         ggml_backend_context_t context;
     };
 
+
+    //
+    // Backend registry
+    //
+
+    typedef ggml_backend_t (*ggml_backend_init_fn)(const char * params, void * user_data);
+
+    void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data);
+
 #ifdef  __cplusplus
 }
 #endif

ggml-backend.h

@@ -7,41 +7,44 @@
 extern "C" {
 #endif
 
+    typedef struct ggml_backend_buffer_type * ggml_backend_buffer_type_t;
+    typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
+    typedef struct ggml_backend * ggml_backend_t;
+    typedef void * ggml_backend_graph_plan_t;
+
     //
     // Backend buffer
     //
 
-    struct ggml_backend_buffer;
+    // buffer type
-    typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
+    GGML_API ggml_backend_buffer_t ggml_backend_buft_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size);
+    GGML_API size_t ggml_backend_buft_get_alignment (ggml_backend_buffer_type_t buft);
+    GGML_API size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor);
+    GGML_API bool ggml_backend_buft_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend);
 
-    // backend buffer functions
+    // buffer
     GGML_API void   ggml_backend_buffer_free          (ggml_backend_buffer_t buffer);
-    GGML_API size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
     GGML_API void * ggml_backend_buffer_get_base      (ggml_backend_buffer_t buffer);
     GGML_API size_t ggml_backend_buffer_get_size      (ggml_backend_buffer_t buffer);
-    GGML_API size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
     GGML_API void   ggml_backend_buffer_init_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-    GGML_API void   ggml_backend_buffer_free_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
+    GGML_API size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API ggml_backend_buffer_type_t ggml_backend_buffer_type(ggml_backend_buffer_t buffer);
 
     //
     // Backend
     //
 
-    struct ggml_backend;
-    typedef struct ggml_backend * ggml_backend_t;
-    typedef void * ggml_backend_graph_plan_t;
-
-    GGML_API ggml_backend_t ggml_get_backend(const struct ggml_tensor * tensor);
 
     GGML_API const char * ggml_backend_name(ggml_backend_t backend);
     GGML_API void         ggml_backend_free(ggml_backend_t backend);
 
+    GGML_API ggml_backend_buffer_type_t ggml_backend_get_default_buffer_type(ggml_backend_t backend);
     GGML_API ggml_backend_buffer_t      ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size);
-
     GGML_API size_t                     ggml_backend_get_alignment(ggml_backend_t backend);
 
-    GGML_API void ggml_backend_tensor_set_async(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+    GGML_API void ggml_backend_tensor_set_async(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-    GGML_API void ggml_backend_tensor_get_async(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+    GGML_API void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
 
     GGML_API void ggml_backend_tensor_set(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
     GGML_API void ggml_backend_tensor_get(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
@@ -57,6 +60,7 @@ extern "C" {
 
     // tensor copy between different backends
     GGML_API void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst);
+    GGML_API void ggml_backend_tensor_copy_async(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst); // automatic fallback to sync copy
 
     //
     // CPU backend
@@ -68,8 +72,23 @@ extern "C" {
     GGML_API void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads);
 
     // Create a backend buffer from an existing pointer
-    GGML_API ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(ggml_backend_t backend_cpu, void * ptr, size_t size);
+    GGML_API ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size);
 
+    GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void);
+
+    //
+    // Backend registry
+    //
+
+    // The backend registry is a registry of all the available backends, and allows initializing backends in a generic way
+
+    GGML_API size_t                     ggml_backend_reg_get_count(void);
+    GGML_API size_t                     ggml_backend_reg_find_by_name(const char * name);
+    GGML_API ggml_backend_t             ggml_backend_reg_init_backend_from_str(const char * backend_str); // str is name[:params]
+    GGML_API const char *               ggml_backend_reg_get_name(size_t i);
+    GGML_API ggml_backend_t             ggml_backend_reg_init_backend(size_t i, const char * params); // params is backend-specific
+    GGML_API ggml_backend_buffer_type_t ggml_backend_reg_get_default_buffer_type(size_t i);
+    GGML_API ggml_backend_buffer_t      ggml_backend_reg_alloc_buffer(size_t i, size_t size);
 
     //
     // Backend scheduler
@@ -131,6 +150,32 @@ extern "C" {
             ggml_backend_sched_t sched,
             struct ggml_cgraph * graph);
 
+
+    //
+    // Utils
+    //
+
+    struct ggml_backend_graph_copy {
+        ggml_backend_buffer_t buffer;
+        struct ggml_context * ctx_allocated;
+        struct ggml_context * ctx_unallocated;
+        struct ggml_cgraph * graph;
+    };
+
+    // Copy a graph to a different backend
+    GGML_API struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, struct ggml_cgraph * graph);
+    GGML_API void                           ggml_backend_graph_copy_free(struct ggml_backend_graph_copy copy);
+
+    typedef bool (*ggml_backend_eval_callback)(int node_index, struct ggml_tensor * t1, struct ggml_tensor * t2, void * user_data);
+
+    // Compare the output of two backends
+    GGML_API void ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data);
+
+    // Tensor initialization
+    GGML_API void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);
+    GGML_API void ggml_backend_view_init(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+
+
 #ifdef  __cplusplus
 }
 #endif

ggml-cuda.h

@@ -49,7 +49,15 @@ GGML_API int    ggml_cuda_get_device_count(void);
 GGML_API void   ggml_cuda_get_device_description(int device, char * description, size_t description_size);
 
 // backend API
-GGML_API ggml_backend_t ggml_backend_cuda_init(void); // TODO: take a list of devices to use
+GGML_API ggml_backend_t ggml_backend_cuda_init(int device);
+
+GGML_API bool ggml_backend_is_cuda(ggml_backend_t backend);
+GGML_API int  ggml_backend_cuda_get_device(ggml_backend_t backend);
+
+GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
+
+// pinned host buffer for use with CPU backend for faster copies between CPU and GPU
+GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
 
 #ifdef  __cplusplus
 }

ggml-impl.h

@@ -232,7 +232,7 @@ bool   ggml_hash_contains      (const struct ggml_hash_set hash_set, struct ggml
 // returns GGML_HASHTABLE_FULL if table is full, otherwise the current index of the key or where it should be inserted
 size_t ggml_hash_find          (const struct ggml_hash_set hash_set, struct ggml_tensor * key);
 
-// returns GGML_HAHSHTABLE_ALREADY_EXISTS if key already exists, index otherwise, asserts if table is full
+// returns GGML_HASHTABLE_ALREADY_EXISTS if key already exists, index otherwise, asserts if table is full
 size_t ggml_hash_insert        (      struct ggml_hash_set hash_set, struct ggml_tensor * key);
 
 // return index, asserts if table is full

ggml-metal.h

@@ -99,6 +99,12 @@ GGML_API ggml_backend_t ggml_backend_metal_init(void);
 GGML_API bool ggml_backend_is_metal(ggml_backend_t backend);
 
 GGML_API void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb);
+GGML_API ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
+
+// helper to check if the device supports a specific family
+// ideally, the user code should be doing these checks
+// ref: https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf
+GGML_API bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family);
 
 #ifdef __cplusplus
 }

ggml-opencl.cpp

@@ -1,20 +1,18 @@
+#include "ggml.h"
 #include "ggml-opencl.h"
 
 #include <array>
 #include <atomic>
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+#include <limits>
 #include <sstream>
 #include <vector>
-#include <limits>
 
 #define CL_TARGET_OPENCL_VERSION 110
 #include <clblast.h>
 
-#include <stdlib.h>
-#include <stdio.h>
-#include <string.h>
-
-#include "ggml.h"
-
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif

ggml-quants.c

@@ -19,7 +19,7 @@
 #ifdef __wasm_simd128__
 #include <wasm_simd128.h>
 #else
-#ifdef __POWER9_VECTOR__
+#if defined(__POWER9_VECTOR__) || defined(__powerpc64__)
 #include <altivec.h>
 #undef bool
 #define bool _Bool
@@ -3114,7 +3114,7 @@ void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * restri
 
     size_t vl = __riscv_vsetvl_e8m1(qk/2);
 
-    // These tempory registers are for masking and shift operations
+    // These temporary registers are for masking and shift operations
     vuint32m2_t vt_1 = __riscv_vid_v_u32m2(vl);
     vuint32m2_t vt_2 = __riscv_vsll_vv_u32m2(__riscv_vmv_v_x_u32m2(1, vl), vt_1, vl);
 
@@ -4757,7 +4757,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
 
             vl = 16;
 
-            // retreive lane to multiply with scale
+            // retrieve lane to multiply with scale
             vint32m2_t aux0_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a0, 0), (scale[0]), vl);
             vint32m2_t aux0_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a0, 1), (scale[1]), vl);
             vint32m2_t aux1_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a1, 0), (scale[2]), vl);

ggml.h

@@ -215,9 +215,9 @@
 #define GGML_QNT_VERSION_FACTOR 1000 // do not change this
 
 #define GGML_MAX_DIMS           4
-#define GGML_MAX_PARAMS         1024
+#define GGML_MAX_PARAMS         2048
 #define GGML_MAX_CONTEXTS       64
-#define GGML_MAX_SRC            6
+#define GGML_MAX_SRC            10
 #define GGML_MAX_NAME           64
 #define GGML_MAX_OP_PARAMS      64
 #define GGML_DEFAULT_N_THREADS  4
@@ -244,11 +244,10 @@
 #define GGML_ASSERT(x) \
     do { \
         if (!(x)) { \
-            fprintf(stderr, "GGML_ASSERT: %s:%d: %s\n", __FILE__, __LINE__, #x); \
-            fflush(stderr); \
             fflush(stdout); \
+            fprintf(stderr, "GGML_ASSERT: %s:%d: %s\n", __FILE__, __LINE__, #x); \
             ggml_print_backtrace(); \
-            exit(1); \
+            abort(); \
         } \
     } while (0)
 
@@ -284,6 +283,20 @@
     const type prefix##3 = (pointer)->array[3]; \
     GGML_UNUSED(prefix##3);
 
+#define GGML_TENSOR_UNARY_OP_LOCALS \
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) \
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb) \
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne) \
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
+
+#define GGML_TENSOR_BINARY_OP_LOCALS \
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) \
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb) \
+    GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne) \
+    GGML_TENSOR_LOCALS(size_t,  nb1, src1, nb) \
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne) \
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
+
 #ifdef  __cplusplus
 extern "C" {
 #endif
@@ -382,6 +395,7 @@ extern "C" {
         GGML_OP_GROUP_NORM,
 
         GGML_OP_MUL_MAT,
+        GGML_OP_MUL_MAT_ID,
         GGML_OP_OUT_PROD,
 
         GGML_OP_SCALE,
@@ -408,8 +422,10 @@ extern "C" {
         GGML_OP_CONV_TRANSPOSE_2D,
         GGML_OP_POOL_1D,
         GGML_OP_POOL_2D,
-
         GGML_OP_UPSCALE, // nearest interpolate
+        GGML_OP_PAD,
+        GGML_OP_ARGSORT,
+        GGML_OP_LEAKY_RELU,
 
         GGML_OP_FLASH_ATTN,
         GGML_OP_FLASH_FF,
@@ -449,7 +465,8 @@ extern "C" {
         GGML_UNARY_OP_GELU,
         GGML_UNARY_OP_GELU_QUICK,
         GGML_UNARY_OP_SILU,
-        GGML_UNARY_OP_LEAKY
+
+        GGML_UNARY_OP_COUNT,
     };
 
     enum ggml_object_type {
@@ -632,6 +649,9 @@ extern "C" {
     GGML_API const char * ggml_op_name  (enum ggml_op   op);
     GGML_API const char * ggml_op_symbol(enum ggml_op   op);
 
+    GGML_API const char * ggml_unary_op_name(enum ggml_unary_op op);
+    GGML_API const char * ggml_op_desc(const struct ggml_tensor * t); // unary or op name
+
     GGML_API size_t  ggml_element_size(const struct ggml_tensor * tensor);
 
     GGML_API bool    ggml_is_quantized(enum ggml_type type);
@@ -774,6 +794,9 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    // dst = a
+    // view(dst, nb1, nb2, nb3, offset) += b
+    // return dst
     GGML_API struct ggml_tensor * ggml_acc(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -938,15 +961,14 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
-    GGML_API struct ggml_tensor * ggml_leaky(
+    GGML_API struct ggml_tensor * ggml_leaky_relu(
             struct ggml_context * ctx,
-            struct ggml_tensor  * a);
+            struct ggml_tensor  * a, float negative_slope, bool inplace);
 
     GGML_API struct ggml_tensor * ggml_relu_inplace(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
-    // TODO: double-check this computation is correct
     GGML_API struct ggml_tensor * ggml_gelu(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
@@ -1028,6 +1050,16 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    // indirect matrix multiplication
+    //  ggml_mul_mat_id(ctx, as, ids, id, b) ~= ggml_mul_mat(as[ids[id]], b)
+    GGML_API struct ggml_tensor * ggml_mul_mat_id(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * const as[],
+            int                   n_as,
+            struct ggml_tensor  * ids,
+            int                   id,
+            struct ggml_tensor  * b);
+
     // A: m columns, n rows,
     // B: p columns, n rows,
     // result is m columns, p rows
@@ -1235,6 +1267,7 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    // supports 3D: a->ne[2] == b->ne[1]
     GGML_API struct ggml_tensor * ggml_get_rows(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -1283,6 +1316,14 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    // fused soft_max(a*scale + mask)
+    // mask is optional
+    GGML_API struct ggml_tensor * ggml_soft_max_ext(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * mask,
+            float                 scale);
+
     GGML_API struct ggml_tensor * ggml_soft_max_back(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -1513,6 +1554,32 @@ extern "C" {
             struct ggml_tensor  * a,
             int                   scale_factor);
 
+    // pad each dimension with zeros: [x, ..., x] -> [x, ..., x, 0, ..., 0]
+    GGML_API struct ggml_tensor * ggml_pad(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                  p0,
+            int                  p1,
+            int                  p2,
+            int                  p3);
+
+    // sort rows
+    enum ggml_sort_order {
+        GGML_SORT_ASC,
+        GGML_SORT_DESC,
+    };
+
+    GGML_API struct ggml_tensor * ggml_argsort(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            enum ggml_sort_order  order);
+
+    // top k elements per row
+    GGML_API struct ggml_tensor * ggml_top_k(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   k);
+
     GGML_API struct ggml_tensor * ggml_flash_attn(
             struct ggml_context * ctx,
             struct ggml_tensor  * q,
@@ -1574,7 +1641,6 @@ extern "C" {
             int                   kh);
 
     // used in sam
-
     GGML_API struct ggml_tensor * ggml_add_rel_pos(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -1749,7 +1815,7 @@ extern "C" {
     GGML_API struct ggml_cgraph * ggml_new_graph         (struct ggml_context * ctx); // size = GGML_DEFAULT_GRAPH_SIZE, grads = false
     GGML_API struct ggml_cgraph * ggml_new_graph_custom  (struct ggml_context * ctx, size_t size, bool grads);
     GGML_API struct ggml_cgraph * ggml_graph_dup         (struct ggml_context * ctx, struct ggml_cgraph * cgraph);
-    GGML_API struct ggml_cgraph * ggml_graph_view        (struct ggml_context * ctx, struct ggml_cgraph * cgraph, int i0, int i1);
+    GGML_API struct ggml_cgraph   ggml_graph_view        (struct ggml_cgraph * cgraph, int i0, int i1);
     GGML_API void                 ggml_graph_cpy         (struct ggml_cgraph * src, struct ggml_cgraph * dst);
     GGML_API void                 ggml_graph_reset       (struct ggml_cgraph * cgraph);  // zero grads
     GGML_API void                 ggml_graph_clear       (struct ggml_cgraph * cgraph);
@@ -2045,6 +2111,7 @@ extern "C" {
     GGML_API double       gguf_get_val_f64 (const struct gguf_context * ctx, int key_id);
     GGML_API bool         gguf_get_val_bool(const struct gguf_context * ctx, int key_id);
     GGML_API const char * gguf_get_val_str (const struct gguf_context * ctx, int key_id);
+    GGML_API const void * gguf_get_val_data(const struct gguf_context * ctx, int key_id);
     GGML_API int          gguf_get_arr_n   (const struct gguf_context * ctx, int key_id);
     GGML_API const void * gguf_get_arr_data(const struct gguf_context * ctx, int key_id);
     GGML_API const char * gguf_get_arr_str (const struct gguf_context * ctx, int key_id, int i);

whisper.cpp

@@ -1063,7 +1063,7 @@ static ggml_backend_t whisper_backend_init(const whisper_context_params & params
 #ifdef GGML_USE_CUBLAS
     if (params.use_gpu && ggml_cublas_loaded()) {
         WHISPER_LOG_INFO("%s: using CUDA backend\n", __func__);
-        backend_gpu = ggml_backend_cuda_init();
+        backend_gpu = ggml_backend_cuda_init(0);
         if (!backend_gpu) {
             WHISPER_LOG_ERROR("%s: ggml_backend_cuda_init() failed\n", __func__);
         }
@@ -1077,6 +1077,10 @@ static ggml_backend_t whisper_backend_init(const whisper_context_params & params
         backend_gpu = ggml_backend_metal_init();
         if (!backend_gpu) {
             WHISPER_LOG_ERROR("%s: ggml_backend_metal_init() failed\n", __func__);
+        } else if (!ggml_backend_metal_supports_family(backend_gpu, 7)) {
+            WHISPER_LOG_ERROR("%s: Metal GPU does not support family 7 - falling back to CPU\n", __func__);
+            ggml_backend_free(backend_gpu);
+            backend_gpu = NULL;
         }
     }
 #endif
@@ -1341,10 +1345,10 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
             model.e_pe = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_audio_state, n_audio_ctx);
 
             model.e_conv_1_w     = ggml_new_tensor_3d(ctx, vtype,         3, n_mels,     n_audio_state);
-            model.e_conv_1_b     = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, 2*n_audio_ctx, n_audio_state);
+            model.e_conv_1_b     = ggml_new_tensor_2d(ctx, GGML_TYPE_F32,         1,     n_audio_state);
 
             model.e_conv_2_w     = ggml_new_tensor_3d(ctx, vtype,         3, n_audio_state, n_audio_state);
-            model.e_conv_2_b     = ggml_new_tensor_2d(ctx, GGML_TYPE_F32,    n_audio_ctx,   n_audio_state);
+            model.e_conv_2_b     = ggml_new_tensor_2d(ctx, GGML_TYPE_F32,                1, n_audio_state);
 
             model.e_ln_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_audio_state);
             model.e_ln_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_audio_state);
@@ -1574,9 +1578,6 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
 
             auto tensor = model.tensors[name.data()];
 
-            const bool is_conv_bias = (name == "encoder.conv1.bias" || name == "encoder.conv2.bias");
-
-            if (!is_conv_bias) {
             if (ggml_nelements(tensor) != nelements) {
                 WHISPER_LOG_ERROR("%s: tensor '%s' has wrong size in model file\n", __func__, name.data());
                 WHISPER_LOG_ERROR("%s: shape: [%d, %d, %d], expected: [%d, %d, %d]\n",
@@ -1597,7 +1598,6 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
                         __func__, name.data(), ggml_nbytes(tensor), nelements*bpe);
                 return false;
             }
-            }
 
             ggml_backend_t backend = wctx.backend;
 
@@ -1607,7 +1607,7 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
 #ifdef GGML_USE_METAL
                 || ggml_backend_is_metal(backend)
 #endif
-                ) && !is_conv_bias) {
+                )) {
                 // for the CPU and Metal backend, we can read directly into the tensor
                 loader->read(loader->context, tensor->data, ggml_nbytes(tensor));
                 BYTESWAP_TENSOR(tensor);
@@ -1615,24 +1615,7 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
                 // read into a temporary buffer first, then copy to device memory
                 read_buf.resize(ggml_nbytes(tensor));
 
-                // we repeat the 2 bias tensors along dim 0:
-                // [1, 512] -> [3000, 512] (conv1.bias)
-                // [1, 512] -> [1500, 512] (conv2.bias)
-                if (is_conv_bias) {
-                    loader->read(loader->context, read_buf.data(), read_buf.size() / tensor->ne[0]);
-
-                    float * data_f32 = (float *) read_buf.data();
-                    for (int64_t y = 0; y < tensor->ne[1]; ++y) {
-                        const int64_t yy = tensor->ne[1] - y - 1;
-                        const float val = data_f32[yy];
-
-                        for (int64_t x = 0; x < tensor->ne[0]; ++x) {
-                            data_f32[yy*tensor->ne[0] + x] = val;
-                        }
-                    }
-                } else {
                 loader->read(loader->context, read_buf.data(), read_buf.size());
-                }
 
                 ggml_backend_tensor_set(tensor, read_buf.data(), 0, ggml_nbytes(tensor));
             }
@@ -1732,20 +1715,12 @@ static struct ggml_cgraph * whisper_build_graph_conv(
         // convolution + gelu
         {
             cur = ggml_conv_1d_ph(ctx0, model.e_conv_1_w, mel, 1, 1);
-            if (n_ctx == hparams.n_audio_ctx) {
             cur = ggml_add(ctx0, cur, model.e_conv_1_b);
-            } else {
-                cur = ggml_add(ctx0, cur, ggml_cont(ctx0, ggml_view_2d(ctx0, model.e_conv_1_b, cur->ne[0], cur->ne[1], model.e_conv_1_b->nb[1], 0)));
-            }
 
             cur = ggml_gelu(ctx0, cur);
 
             cur = ggml_conv_1d_ph(ctx0, model.e_conv_2_w, cur, 2, 1);
-            if (n_ctx == hparams.n_audio_ctx) {
             cur = ggml_add(ctx0, cur, model.e_conv_2_b);
-            } else {
-                cur = ggml_add(ctx0, cur, ggml_cont(ctx0, ggml_view_2d(ctx0, model.e_conv_2_b, cur->ne[0], cur->ne[1], model.e_conv_2_b->nb[1], 0)));
-            }
 
             cur = ggml_gelu(ctx0, cur);
         }
@@ -3593,6 +3568,17 @@ const char * whisper_lang_str(int id) {
     return nullptr;
 }
 
+const char * whisper_lang_str_full(int id) {
+   for (const auto & kv : g_lang) {
+        if (kv.second.first == id) {
+            return kv.second.second.c_str();
+        }
+    }
+
+    WHISPER_LOG_ERROR("%s: unknown language id %d\n", __func__, id);
+    return nullptr;
+}
+
 int whisper_lang_auto_detect_with_state(
         struct whisper_context * ctx,
           struct whisper_state * state,
@@ -5042,6 +5028,7 @@ int whisper_full_with_state(
     // basically don't process anything that is less than 1.0s
     // see issue #39: https://github.com/ggerganov/whisper.cpp/issues/39
     if (seek_end < seek_start + (params.speed_up ? 50 : 100)) {
+        WHISPER_PRINT_DEBUG("%s: input is too short - %d ms < 1000 ms\n", __func__, (seek_end - seek_start)*10);
         return 0;
     }
 
@@ -5180,7 +5167,7 @@ int whisper_full_with_state(
                 ctx, state, progress_cur, params.progress_callback_user_data);
         }
 
-        // of only 1 second left, then stop
+        // if only 1 second left, then stop
         if (seek + 100 >= seek_end) {
             break;
         }
@@ -5469,6 +5456,7 @@ int whisper_full_with_state(
 
                             // do not allow to go back in time
                             if (has_ts && seek_delta > seek_delta_new && result_len < i) {
+                                WHISPER_PRINT_DEBUG("%s: decoder %d: failed due to seek_delta (%d > %d)\n", __func__, j, seek_delta, seek_delta_new);
                                 failed = true; // TODO: maybe this is not a failure ?
                                 continue;
                             }
@@ -5497,6 +5485,7 @@ int whisper_full_with_state(
                                 if (seek + seek_delta + 100 >= seek_end) {
                                     result_len = i + 1;
                                 } else {
+                                    WHISPER_PRINT_DEBUG("%s: decoder %d failed (result_len = 0)\n", __func__, j);
                                     failed = true;
                                     continue;
                                 }
@@ -5507,6 +5496,7 @@ int whisper_full_with_state(
                                 seek_delta = 100*WHISPER_CHUNK_SIZE;
                             }
 
+                            WHISPER_PRINT_DEBUG("%s: decoder %d completed\n", __func__, j);
                             completed = true;
                             continue;
                         }
@@ -5522,6 +5512,7 @@ int whisper_full_with_state(
                     // sometimes, the decoding can get stuck in a repetition loop
                     // this is an attempt to mitigate such cases - we flag the decoding as failed and use a fallback strategy
                     if (i == n_max - 1 && (result_len == 0 || seek_delta < 100*WHISPER_CHUNK_SIZE/2)) {
+                        WHISPER_PRINT_DEBUG("%s: decoder %d: failed due to repetition loop\n", __func__, j);
                         failed = true;
                         continue;
                     }
@@ -5665,20 +5656,20 @@ int whisper_full_with_state(
                 WHISPER_PRINT_DEBUG("%s: best decoder = %d\n", __func__, best_decoder_id);
             }
 
+            bool success = true;
+
             // was the decoding successful for the current temperature?
             // do fallback only if:
             // - we are not at the last temperature
-            // - we are not at the end of the audio (3 sec)
+            if (it != (int) temperatures.size() - 1) {
-            if (it != (int) temperatures.size() - 1 &&
-                seek_end - seek > 10*WHISPER_CHUNK_SIZE) {
-                bool success = true;
-
                 const auto & decoder = state->decoders[best_decoder_id];
 
                 if (decoder.failed || decoder.sequence.avg_logprobs < params.logprob_thold) {
+                    WHISPER_PRINT_DEBUG("%s: failed due to avg_logprobs %8.5f < %8.5f\n", __func__, decoder.sequence.avg_logprobs, params.logprob_thold);
                     success = false;
                     state->n_fail_p++;
                 }
+            }
 
             if (success) {
                 //for (auto & token : ctx->decoders[best_decoder_id].sequence.tokens) {
@@ -5687,7 +5678,6 @@ int whisper_full_with_state(
 
                 break;
             }
-            }
 
             WHISPER_PRINT_DEBUG("\n%s: failed to decode with temperature = %.2f\n", __func__, t_cur);
         }
@@ -6138,7 +6128,7 @@ WHISPER_API const char * whisper_bench_memcpy_str(int n_threads) {
 
     // multi-thread
 
-    for (uint32_t n_threads = 1; n_threads <= std::thread::hardware_concurrency(); n_threads++) {
+    for (uint32_t k = 1; k <= n_threads; k++) {
         char * src = (char *) malloc(size);
         char * dst = (char *) malloc(size);
 
@@ -6149,8 +6139,8 @@ WHISPER_API const char * whisper_bench_memcpy_str(int n_threads) {
         double tsum = 0.0;
 
         auto helper = [&](int th) {
-            const int64_t i0 = (th + 0)*size/n_threads;
+            const int64_t i0 = (th + 0)*size/k;
-            const int64_t i1 = (th + 1)*size/n_threads;
+            const int64_t i1 = (th + 1)*size/k;
 
             for (size_t i = 0; i < n; i++) {
                 memcpy(dst + i0, src + i0, i1 - i0);
@@ -6161,14 +6151,14 @@ WHISPER_API const char * whisper_bench_memcpy_str(int n_threads) {
 
         const int64_t t0 = ggml_time_us();
 
-        std::vector<std::thread> threads(n_threads - 1);
+        std::vector<std::thread> threads(k - 1);
-        for (uint32_t th = 0; th < n_threads - 1; ++th) {
+        for (uint32_t th = 0; th < k - 1; ++th) {
             threads[th] = std::thread(helper, th);
         }
 
-        helper(n_threads - 1);
+        helper(k - 1);
 
-        for (uint32_t th = 0; th < n_threads - 1; ++th) {
+        for (uint32_t th = 0; th < k - 1; ++th) {
             threads[th].join();
         }
 
@@ -6176,7 +6166,7 @@ WHISPER_API const char * whisper_bench_memcpy_str(int n_threads) {
 
         tsum += (t1 - t0)*1e-6;
 
-        snprintf(strbuf, sizeof(strbuf), "memcpy: %7.2f GB/s (%2d thread)\n", (double) (n*size)/(tsum*1e9), n_threads);
+        snprintf(strbuf, sizeof(strbuf), "memcpy: %7.2f GB/s (%2d thread)\n", (double) (n*size)/(tsum*1e9), k);
         s += strbuf;
 
         // needed to prevent the compiler from optimizing the memcpy away

whisper.h

@@ -315,6 +315,9 @@ extern "C" {
     // Return the short string of the specified language id (e.g. 2 -> "de"), returns nullptr if not found
     WHISPER_API const char * whisper_lang_str(int id);
 
+    // Return the short string of the specified language name (e.g. 2 -> "german"), returns nullptr if not found
+    WHISPER_API const char * whisper_lang_str_full(int id);
+
     // 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 top language id or negative on failure