rebase on master after whisper_state changes

* fixed blocking code on node addon

* modify the example to run async

* format

* added logic to see the whisper output

* added logic to see the whisper output

* removed extra function for more clean example

.gitignore

@@ -1,5 +1,7 @@
 *.o
 *.a
+*.mlmodel
+*.mlmodelc
 .cache/
 .vs/
 .vscode/

CMakeLists.txt

@@ -1,6 +1,6 @@
 cmake_minimum_required (VERSION 3.0)
 
-project(whisper.cpp VERSION 1.2.0)
+project(whisper.cpp VERSION 1.2.1)
 
 # Add path to modules
 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
@@ -54,6 +54,8 @@ if (APPLE)
     option(WHISPER_NO_AVX              "whisper: disable AVX" OFF)
     option(WHISPER_NO_AVX2             "whisper: disable AVX2" OFF)
     option(WHISPER_NO_FMA              "whisper: disable FMA" OFF)
+
+    option(WHISPER_COREML              "whisper: enable Core ML framework" OFF)
 else()
     option(WHISPER_SUPPORT_OPENBLAS    "whisper: support for OpenBLAS" OFF)
 endif()
@@ -86,16 +88,33 @@ endif()
 
 find_package(Threads REQUIRED)
 
-# on APPLE - include Accelerate framework
-if (APPLE AND NOT WHISPER_NO_ACCELERATE)
-    find_library(ACCELERATE_FRAMEWORK Accelerate)
-    if (ACCELERATE_FRAMEWORK)
-        message(STATUS "Accelerate framework found")
+# on APPLE
+if (APPLE)
+    # include Accelerate framework
+    if (NOT WHISPER_NO_ACCELERATE)
+        find_library(ACCELERATE_FRAMEWORK Accelerate)
 
-        set(WHISPER_EXTRA_LIBS  ${WHISPER_EXTRA_LIBS}  ${ACCELERATE_FRAMEWORK})
-        set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_ACCELERATE)
-    else()
-        message(WARNING "Accelerate framework not found")
+        if (ACCELERATE_FRAMEWORK)
+            message(STATUS "Accelerate framework found")
+
+            set(WHISPER_EXTRA_LIBS  ${WHISPER_EXTRA_LIBS}  ${ACCELERATE_FRAMEWORK})
+            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_ACCELERATE)
+        else()
+            message(WARNING "Accelerate framework not found")
+        endif()
+    endif()
+
+    if (WHISPER_COREML)
+        find_library(FOUNDATION_FRAMEWORK Foundation)
+        find_library(COREML_FRAMEWORK CoreML)
+
+        if (COREML_FRAMEWORK)
+            message(STATUS "CoreML framework found")
+
+            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DWHISPER_USE_COREML)
+        else()
+            message(WARNING "CoreML framework not found")
+        endif()
     endif()
 endif()
 
@@ -172,7 +191,9 @@ else()
             if(NOT WHISPER_NO_FMA)
                 set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mfma")
             endif()
-            set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mf16c")
+            if(NOT WHISPER_NO_F16C)
+                set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mf16c")
+            endif()
         endif()
     endif()
 endif()
@@ -181,6 +202,33 @@ if (WHISPER_PERF)
     set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_PERF)
 endif()
 
+#
+# whisper.coreml - Core ML support
+#
+
+if (WHISPER_COREML)
+    set(TARGET whisper.coreml)
+
+    add_library(${TARGET}
+        coreml/whisper-encoder.h
+        coreml/whisper-encoder.mm
+        coreml/whisper-encoder-impl.h
+        coreml/whisper-encoder-impl.m
+        )
+
+    include(DefaultTargetOptions)
+
+    target_include_directories(${TARGET} PUBLIC
+        .
+        )
+
+    target_link_libraries(${TARGET} PRIVATE ${FOUNDATION_FRAMEWORK} ${COREML_FRAMEWORK})
+
+    set_target_properties(${TARGET} PROPERTIES
+        COMPILE_FLAGS "-fobjc-arc"
+        )
+endif()
+
 #
 # whisper - this is the main library of the project
 #
@@ -200,6 +248,10 @@ target_include_directories(${TARGET} PUBLIC
     .
     )
 
+if (WHISPER_COREML)
+    target_link_libraries(${TARGET} PRIVATE whisper.coreml)
+endif()
+
 if (MSVC)
     target_link_libraries(${TARGET} PRIVATE ${WHISPER_EXTRA_LIBS} ${CMAKE_THREAD_LIBS_INIT})
 

Makefile

@@ -30,10 +30,16 @@ endif
 # Compile flags
 #
 
-CFLAGS   = -I.              -O3 -std=c11   -fPIC
-CXXFLAGS = -I. -I./examples -O3 -std=c++11 -fPIC
+CFLAGS   = -I.              -O3 -DNDEBUG -std=c11   -fPIC
+CXXFLAGS = -I. -I./examples -O3 -DNDEBUG -std=c++11 -fPIC
 LDFLAGS  =
 
+# ref: https://github.com/ggerganov/whisper.cpp/issues/37
+ifneq ($(wildcard /usr/include/musl/*),)
+	CFLAGS   += -D_POSIX_SOURCE -D_GNU_SOURCE
+	CXXFLAGS += -D_POSIX_SOURCE -D_GNU_SOURCE
+endif
+
 # OS specific
 # TODO: support Windows
 ifeq ($(UNAME_S),Linux)
@@ -132,6 +138,10 @@ ifndef WHISPER_NO_ACCELERATE
 		LDFLAGS += -framework Accelerate
 	endif
 endif
+ifdef WHISPER_COREML
+	CXXFLAGS += -DWHISPER_USE_COREML
+	LDFLAGS  += -framework Foundation -framework CoreML
+endif
 ifdef WHISPER_OPENBLAS
 	CFLAGS  += -DGGML_USE_OPENBLAS -I/usr/local/include/openblas
 	LDFLAGS += -lopenblas
@@ -141,6 +151,8 @@ ifdef WHISPER_GPROF
 	CXXFLAGS += -pg
 endif
 ifneq ($(filter aarch64%,$(UNAME_M)),)
+	CFLAGS += -mcpu=native
+	CXXFLAGS += -mcpu=native
 endif
 ifneq ($(filter armv6%,$(UNAME_M)),)
 	# Raspberry Pi 1, 2, 3
@@ -182,11 +194,23 @@ ggml.o: ggml.c ggml.h
 whisper.o: whisper.cpp whisper.h
 	$(CXX) $(CXXFLAGS) -c whisper.cpp -o whisper.o
 
-libwhisper.a: ggml.o whisper.o
-	$(AR) rcs libwhisper.a ggml.o whisper.o
+ifndef WHISPER_COREML
+WHISPER_OBJ = whisper.o
+else
+whisper-encoder.o: coreml/whisper-encoder.mm coreml/whisper-encoder.h
+	$(CXX) -O3 -I . -c coreml/whisper-encoder.mm -o whisper-encoder.o
 
-libwhisper.so: ggml.o whisper.o
-	$(CXX) $(CXXFLAGS) -shared -o libwhisper.so ggml.o whisper.o $(LDFLAGS)
+whisper-encoder-impl.o: coreml/whisper-encoder-impl.m coreml/whisper-encoder-impl.h
+	$(CXX) -O3 -I . -fobjc-arc -c coreml/whisper-encoder-impl.m -o whisper-encoder-impl.o
+
+WHISPER_OBJ = whisper.o whisper-encoder.o whisper-encoder-impl.o
+endif
+
+libwhisper.a: ggml.o $(WHISPER_OBJ)
+	$(AR) rcs libwhisper.a ggml.o $(WHISPER_OBJ)
+
+libwhisper.so: ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) -shared -o libwhisper.so ggml.o $(WHISPER_OBJ) $(LDFLAGS)
 
 clean:
 	rm -f *.o main stream command talk bench libwhisper.a libwhisper.so
@@ -200,21 +224,21 @@ CC_SDL=`sdl2-config --cflags --libs`
 SRC_COMMON = examples/common.cpp
 SRC_COMMON_SDL = examples/common-sdl.cpp
 
-main: examples/main/main.cpp $(SRC_COMMON) ggml.o whisper.o
-	$(CXX) $(CXXFLAGS) examples/main/main.cpp $(SRC_COMMON) ggml.o whisper.o -o main $(LDFLAGS)
+main: examples/main/main.cpp $(SRC_COMMON) ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/main/main.cpp $(SRC_COMMON) ggml.o $(WHISPER_OBJ) -o main $(LDFLAGS)
 	./main -h
 
-stream: examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o whisper.o
-	$(CXX) $(CXXFLAGS) examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o whisper.o -o stream $(CC_SDL) $(LDFLAGS)
+stream: examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ) -o stream $(CC_SDL) $(LDFLAGS)
 
-command: examples/command/command.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o whisper.o
-	$(CXX) $(CXXFLAGS) examples/command/command.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o whisper.o -o command $(CC_SDL) $(LDFLAGS)
+command: examples/command/command.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/command/command.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ) -o command $(CC_SDL) $(LDFLAGS)
 
-talk: examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o whisper.o
-	$(CXX) $(CXXFLAGS) examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o whisper.o -o talk $(CC_SDL) $(LDFLAGS)
+talk: examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ) -o talk $(CC_SDL) $(LDFLAGS)
 
-bench: examples/bench/bench.cpp ggml.o whisper.o
-	$(CXX) $(CXXFLAGS) examples/bench/bench.cpp ggml.o whisper.o -o bench $(LDFLAGS)
+bench: examples/bench/bench.cpp ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/bench/bench.cpp ggml.o $(WHISPER_OBJ) -o bench $(LDFLAGS)
 
 #
 # Audio samples

README.md

@@ -4,7 +4,7 @@
 [![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: [v1.2.0](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.2.0) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)
+Stable: [v1.2.1](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.2.1) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)
 
 High-performance inference of [OpenAI's Whisper](https://github.com/openai/whisper) automatic speech recognition (ASR) model:
 
@@ -433,6 +433,19 @@ https://user-images.githubusercontent.com/1991296/199337538-b7b0c7a3-2753-4a88-a
 
 ---
 
+## Video comparison of different models
+
+Use the [extra/bench-wts.sh](https://github.com/ggerganov/whisper.cpp/blob/master/extra/bench-wts.sh) script to generate a video in the following format:
+
+```java
+./extra/bench-wts.sh samples/jfk.wav
+ffplay ./samples/jfk.wav.all.mp4
+```
+
+https://user-images.githubusercontent.com/1991296/223206245-2d36d903-cf8e-4f09-8c3b-eb9f9c39d6fc.mp4
+
+---
+
 ## Benchmarks
 
 In order to have an objective comparison of the performance of the inference across different system configurations,
@@ -453,7 +466,7 @@ The original models are converted to a custom binary format. This allows to pack
 You can download the converted models using the [models/download-ggml-model.sh](models/download-ggml-model.sh) script
 or manually from here:
 
-- https://huggingface.co/datasets/ggerganov/whisper.cpp
+- https://huggingface.co/ggerganov/whisper.cpp
 - https://ggml.ggerganov.com
 
 For more details, see the conversion script [models/convert-pt-to-ggml.py](models/convert-pt-to-ggml.py) or the README
@@ -463,6 +476,7 @@ in [models](models).
 
 - [X] Rust: [tazz4843/whisper-rs](https://github.com/tazz4843/whisper-rs) | [#310](https://github.com/ggerganov/whisper.cpp/discussions/310)
 - [X] Javascript: [bindings/javascript](bindings/javascript) | [#309](https://github.com/ggerganov/whisper.cpp/discussions/309)
+  - React Native (iOS / Android): [whisper.rn](https://github.com/mybigday/whisper.rn)
 - [X] Go: [bindings/go](bindings/go) | [#312](https://github.com/ggerganov/whisper.cpp/discussions/312)
 - [X] Ruby: [bindings/ruby](bindings/ruby) | [#507](https://github.com/ggerganov/whisper.cpp/discussions/507)
 - [X] Objective-C / Swift: [ggerganov/whisper.spm](https://github.com/ggerganov/whisper.spm) | [#313](https://github.com/ggerganov/whisper.cpp/discussions/313)
@@ -471,6 +485,8 @@ in [models](models).
   - [NickDarvey/whisper](https://github.com/NickDarvey/whisper)
 - [X] Python: | [#9](https://github.com/ggerganov/whisper.cpp/issues/9)
   - [stlukey/whispercpp.py](https://github.com/stlukey/whispercpp.py) (Cython)
+  - [aarnphm/whispercpp](https://github.com/aarnphm/whispercpp) (Pybind11)
+- [X] R: [bnosac/audio.whisper](https://github.com/bnosac/audio.whisper)
 
 ## Examples
 

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

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

bindings/go/pkg/whisper/model.go

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

bindings/go/whisper.go

@@ -20,7 +20,7 @@ extern bool callEncoderBegin(void* user_data);
 // Text segment callback
 // Called on every newly generated text segment
 // Use the whisper_full_...() functions to obtain the text segments
-static void whisper_new_segment_cb(struct whisper_context* ctx, int n_new, void* user_data) {
+static void whisper_new_segment_cb(struct whisper_context* ctx, struct whisper_state* state, int n_new, void* user_data) {
     if(user_data != NULL && ctx != NULL) {
         callNewSegment(user_data, n_new);
     }
@@ -29,7 +29,7 @@ static void whisper_new_segment_cb(struct whisper_context* ctx, int n_new, void*
 // Encoder begin callback
 // If not NULL, called before the encoder starts
 // If it returns false, the computation is aborted
-static bool whisper_encoder_begin_cb(struct whisper_context* ctx, void* user_data) {
+static bool whisper_encoder_begin_cb(struct whisper_context* ctx, struct whisper_state* state, void* user_data) {
     if(user_data != NULL && ctx != NULL) {
         return callEncoderBegin(user_data);
     }

bindings/javascript/package.json

@@ -1,6 +1,6 @@
 {
   "name": "whisper.cpp",
-  "version": "1.2.0",
+  "version": "1.2.1",
   "description": "Whisper speech recognition",
   "main": "whisper.js",
   "scripts": {

bindings/ruby/ext/ruby_whisper.cpp

@@ -199,7 +199,7 @@ static VALUE ruby_whisper_transcribe(int argc, VALUE *argv, VALUE self) {
   {
     static bool is_aborted = false; // NOTE: this should be atomic to avoid data race
 
-    rwp->params.encoder_begin_callback = [](struct whisper_context * /*ctx*/, void * user_data) {
+    rwp->params.encoder_begin_callback = [](struct whisper_context * /*ctx*/, struct whisper_state * /*state*/, void * user_data) {
       bool is_aborted = *(bool*)user_data;
       return !is_aborted;
     };

coreml/whisper-encoder-impl.h

@@ -0,0 +1,142 @@
+//
+// CoremlEncoder.h
+//
+// This file was automatically generated and should not be edited.
+//
+
+#import <Foundation/Foundation.h>
+#import <CoreML/CoreML.h>
+#include <stdint.h>
+#include <os/log.h>
+
+NS_ASSUME_NONNULL_BEGIN
+
+
+/// Model Prediction Input Type
+API_AVAILABLE(macos(10.15), ios(13.0), watchos(6.0), tvos(13.0)) __attribute__((visibility("hidden")))
+@interface CoremlEncoderInput : NSObject<MLFeatureProvider>
+
+/// melSegment as 1 × 80 × 3000 3-dimensional array of floats
+@property (readwrite, nonatomic, strong) MLMultiArray * melSegment;
+- (instancetype)init NS_UNAVAILABLE;
+- (instancetype)initWithMelSegment:(MLMultiArray *)melSegment NS_DESIGNATED_INITIALIZER;
+
+@end
+
+
+/// Model Prediction Output Type
+API_AVAILABLE(macos(10.15), ios(13.0), watchos(6.0), tvos(13.0)) __attribute__((visibility("hidden")))
+@interface CoremlEncoderOutput : NSObject<MLFeatureProvider>
+
+/// output as multidimensional array of floats
+@property (readwrite, nonatomic, strong) MLMultiArray * output;
+- (instancetype)init NS_UNAVAILABLE;
+- (instancetype)initWithOutput:(MLMultiArray *)output NS_DESIGNATED_INITIALIZER;
+
+@end
+
+
+/// Class for model loading and prediction
+API_AVAILABLE(macos(10.15), ios(13.0), watchos(6.0), tvos(13.0)) __attribute__((visibility("hidden")))
+@interface CoremlEncoder : NSObject
+@property (readonly, nonatomic, nullable) MLModel * model;
+
+/**
+    URL of the underlying .mlmodelc directory.
+*/
++ (nullable NSURL *)URLOfModelInThisBundle;
+
+/**
+    Initialize CoremlEncoder instance from an existing MLModel object.
+
+    Usually the application does not use this initializer unless it makes a subclass of CoremlEncoder.
+    Such application may want to use `-[MLModel initWithContentsOfURL:configuration:error:]` and `+URLOfModelInThisBundle` to create a MLModel object to pass-in.
+*/
+- (instancetype)initWithMLModel:(MLModel *)model NS_DESIGNATED_INITIALIZER;
+
+/**
+    Initialize CoremlEncoder instance with the model in this bundle.
+*/
+- (nullable instancetype)init;
+
+/**
+    Initialize CoremlEncoder instance with the model in this bundle.
+
+    @param configuration The model configuration object
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+*/
+- (nullable instancetype)initWithConfiguration:(MLModelConfiguration *)configuration error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Initialize CoremlEncoder instance from the model URL.
+
+    @param modelURL URL to the .mlmodelc directory for CoremlEncoder.
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+*/
+- (nullable instancetype)initWithContentsOfURL:(NSURL *)modelURL error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Initialize CoremlEncoder instance from the model URL.
+
+    @param modelURL URL to the .mlmodelc directory for CoremlEncoder.
+    @param configuration The model configuration object
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+*/
+- (nullable instancetype)initWithContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Construct CoremlEncoder instance asynchronously with configuration.
+    Model loading may take time when the model content is not immediately available (e.g. encrypted model). Use this factory method especially when the caller is on the main thread.
+
+    @param configuration The model configuration
+    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid CoremlEncoder instance or NSError object.
+*/
++ (void)loadWithConfiguration:(MLModelConfiguration *)configuration completionHandler:(void (^)(CoremlEncoder * _Nullable model, NSError * _Nullable error))handler API_AVAILABLE(macos(11.0), ios(14.0), watchos(7.0), tvos(14.0)) __attribute__((visibility("hidden")));
+
+/**
+    Construct CoremlEncoder instance asynchronously with URL of .mlmodelc directory and optional configuration.
+
+    Model loading may take time when the model content is not immediately available (e.g. encrypted model). Use this factory method especially when the caller is on the main thread.
+
+    @param modelURL The model URL.
+    @param configuration The model configuration
+    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid CoremlEncoder instance or NSError object.
+*/
++ (void)loadContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration completionHandler:(void (^)(CoremlEncoder * _Nullable model, NSError * _Nullable error))handler API_AVAILABLE(macos(11.0), ios(14.0), watchos(7.0), tvos(14.0)) __attribute__((visibility("hidden")));
+
+/**
+    Make a prediction using the standard interface
+    @param input an instance of CoremlEncoderInput to predict from
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+    @return the prediction as CoremlEncoderOutput
+*/
+- (nullable CoremlEncoderOutput *)predictionFromFeatures:(CoremlEncoderInput *)input error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Make a prediction using the standard interface
+    @param input an instance of CoremlEncoderInput to predict from
+    @param options prediction options
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+    @return the prediction as CoremlEncoderOutput
+*/
+- (nullable CoremlEncoderOutput *)predictionFromFeatures:(CoremlEncoderInput *)input options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Make a prediction using the convenience interface
+    @param melSegment as 1 × 80 × 3000 3-dimensional array of floats:
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+    @return the prediction as CoremlEncoderOutput
+*/
+- (nullable CoremlEncoderOutput *)predictionFromMelSegment:(MLMultiArray *)melSegment error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Batch prediction
+    @param inputArray array of CoremlEncoderInput instances to obtain predictions from
+    @param options prediction options
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+    @return the predictions as NSArray<CoremlEncoderOutput *>
+*/
+- (nullable NSArray<CoremlEncoderOutput *> *)predictionsFromInputs:(NSArray<CoremlEncoderInput*> *)inputArray options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+@end
+
+NS_ASSUME_NONNULL_END

coreml/whisper-encoder-impl.m

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

coreml/whisper-encoder.h

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

coreml/whisper-encoder.mm

@@ -0,0 +1,61 @@
+#import "coreml/whisper-encoder.h"
+#import "coreml/whisper-encoder-impl.h"
+
+#import <CoreML/CoreML.h>
+
+#include <stdlib.h>
+
+#if __cplusplus
+extern "C" {
+#endif
+
+struct whisper_coreml_context {
+    const void * data;
+};
+
+struct whisper_coreml_context * whisper_coreml_init(const char * path_model) {
+    NSString * path_model_str = [[NSString alloc] initWithUTF8String:path_model];
+
+    NSURL * url_model = [NSURL fileURLWithPath: path_model_str];
+
+    const void * data = CFBridgingRetain([[CoremlEncoder alloc] initWithContentsOfURL:url_model error:nil]);
+
+    if (data == NULL) {
+        return NULL;
+    }
+
+    whisper_coreml_context * ctx = new whisper_coreml_context;
+
+    ctx->data = data;
+
+    return ctx;
+}
+
+void whisper_coreml_free(struct whisper_coreml_context * ctx) {
+    CFRelease(ctx->data);
+    delete ctx;
+}
+
+void whisper_coreml_encode(
+        const whisper_coreml_context * ctx,
+                               float * mel,
+                               float * out) {
+    MLMultiArray * inMultiArray = [
+        [MLMultiArray alloc] initWithDataPointer: mel
+                                           shape: @[@1, @80, @3000]
+                                        dataType: MLMultiArrayDataTypeFloat32
+                                         strides: @[@(240000), @(3000), @1]
+                                     deallocator: nil
+                                           error: nil
+    ];
+
+    CoremlEncoderOutput * outCoreML = [(__bridge id) ctx->data predictionFromMelSegment:inMultiArray error:nil];
+
+    MLMultiArray * outMA = outCoreML.output;
+
+    memcpy(out, outMA.dataPointer, outMA.count * sizeof(float));
+}
+
+#if __cplusplus
+}
+#endif

examples/addon.node/addon.cpp

@@ -72,7 +72,7 @@ int timestamp_to_sample(int64_t t, int n_samples) {
     return std::max(0, std::min((int) n_samples - 1, (int) ((t*WHISPER_SAMPLE_RATE)/100)));
 }
 
-void whisper_print_segment_callback(struct whisper_context * ctx, int n_new, void * user_data) {
+void whisper_print_segment_callback(struct whisper_context * ctx, struct whisper_state * state, int n_new, void * user_data) {
     const auto & params  = *((whisper_print_user_data *) user_data)->params;
     const auto & pcmf32s = *((whisper_print_user_data *) user_data)->pcmf32s;
 
@@ -260,7 +260,7 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
             {
                 static bool is_aborted = false; // NOTE: this should be atomic to avoid data race
 
-                wparams.encoder_begin_callback = [](struct whisper_context * /*ctx*/, void * user_data) {
+                wparams.encoder_begin_callback = [](struct whisper_context * /*ctx*/, struct whisper_state * /*state*/, void * user_data) {
                     bool is_aborted = *(bool*)user_data;
                     return !is_aborted;
                 };
@@ -292,51 +292,64 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
     return 0;
 }
 
-Napi::Object whisper(const Napi::CallbackInfo& info) {
-    Napi::Env env = info.Env();
-    if (info.Length() <= 0 || !info[0].IsObject()) {
-        Napi::TypeError::New(env, "object expected").ThrowAsJavaScriptException();
-    }
-    whisper_params params;
-    std::vector<std::vector<std::string>> result;
+class Worker : public Napi::AsyncWorker {
+ public:
+  Worker(Napi::Function& callback, whisper_params params)
+      : Napi::AsyncWorker(callback), params(params) {}
 
-    Napi::Object whisper_params = info[0].As<Napi::Object>();
-    std::string language = whisper_params.Get("language").As<Napi::String>();
-    std::string model = whisper_params.Get("model").As<Napi::String>();
-    std::string input = whisper_params.Get("fname_inp").As<Napi::String>();
-
-    params.language = language;
-    params.model = model;
-    params.fname_inp.emplace_back(input);
-
-    // run model
+  void Execute() override {
     run(params, result);
+  }
 
-    fprintf(stderr, "RESULT:\n");
-    for (auto sentence:result) {
-        fprintf(stderr, "t0: %s, t1: %s, content: %s \n",
-                sentence[0].c_str(), sentence[1].c_str(), sentence[2].c_str());
-    }
-
-    Napi::Object res = Napi::Array::New(env, result.size());
+  void OnOK() override {
+    Napi::HandleScope scope(Env());
+    Napi::Object res = Napi::Array::New(Env(), result.size());
     for (uint64_t i = 0; i < result.size(); ++i) {
-        Napi::Object tmp = Napi::Array::New(env, 3);
-        for (uint64_t j = 0; j < 3; ++j) {
-            tmp[j] = Napi::String::New(env, result[i][j]);
-        }
-        res[i] = tmp;
+      Napi::Object tmp = Napi::Array::New(Env(), 3);
+      for (uint64_t j = 0; j < 3; ++j) {
+        tmp[j] = Napi::String::New(Env(), result[i][j]);
+      }
+      res[i] = tmp;
     }
+    Callback().Call({Env().Null(), res});
+  }
 
-    return res;
+ private:
+  whisper_params params;
+  std::vector<std::vector<std::string>> result;
+};
+
+
+
+Napi::Value whisper(const Napi::CallbackInfo& info) {
+  Napi::Env env = info.Env();
+  if (info.Length() <= 0 || !info[0].IsObject()) {
+    Napi::TypeError::New(env, "object expected").ThrowAsJavaScriptException();
+  }
+  whisper_params params;
+
+  Napi::Object whisper_params = info[0].As<Napi::Object>();
+  std::string language = whisper_params.Get("language").As<Napi::String>();
+  std::string model = whisper_params.Get("model").As<Napi::String>();
+  std::string input = whisper_params.Get("fname_inp").As<Napi::String>();
+
+  params.language = language;
+  params.model = model;
+  params.fname_inp.emplace_back(input);
+
+  Napi::Function callback = info[1].As<Napi::Function>();
+  Worker* worker = new Worker(callback, params);
+  worker->Queue();
+  return env.Undefined();
 }
 
 
 Napi::Object Init(Napi::Env env, Napi::Object exports) {
-    exports.Set(
-            Napi::String::New(env, "whisper"),
-            Napi::Function::New(env, whisper)
-    );
-    return exports;
+  exports.Set(
+      Napi::String::New(env, "whisper"),
+      Napi::Function::New(env, whisper)
+  );
+  return exports;
 }
 
 NODE_API_MODULE(whisper, Init);

examples/addon.node/index.js

@@ -1,27 +1,36 @@
-const path = require('path');
-const { whisper } = require(path.join(__dirname, '../../build/Release/whisper-addon'));
+const path = require("path");
+const { whisper } = require(path.join(
+  __dirname,
+  "../../build/Release/whisper-addon"
+));
+const { promisify } = require("util");
+
+const whisperAsync = promisify(whisper);
 
 const whisperParams = {
-    language: 'en',
-    model: path.join(__dirname, '../../models/ggml-base.en.bin'),
-    fname_inp: '',
+  language: "en",
+  model: path.join(__dirname, "../../models/ggml-base.en.bin"),
+  fname_inp: "../../samples/jfk.wav",
 };
 
 const arguments = process.argv.slice(2);
 const params = Object.fromEntries(
-    arguments.reduce((pre, item) => {
-        if (item.startsWith("--")) {
-            return [...pre, item.slice(2).split("=")];
-        }
-        return pre;
-    }, []),
+  arguments.reduce((pre, item) => {
+    if (item.startsWith("--")) {
+      return [...pre, item.slice(2).split("=")];
+    }
+    return pre;
+  }, [])
 );
 
 for (const key in params) {
-    if (whisperParams.hasOwnProperty(key)) {
-        whisperParams[key] = params[key];
-    }
+  if (whisperParams.hasOwnProperty(key)) {
+    whisperParams[key] = params[key];
+  }
 }
 
-console.log('whisperParams =', whisperParams);
-console.log(whisper(whisperParams));
+console.log("whisperParams =", whisperParams);
+
+whisperAsync(whisperParams).then((result) => {
+  console.log(`Result from whisper: ${result}`);
+});

examples/helpers.js

@@ -145,15 +145,7 @@ function loadRemote(url, dst, size_mb, cbProgress, cbReady, cbCancel, cbPrint) {
                             var db = event.target.result;
                             var tx = db.transaction(['models'], 'readwrite');
                             var os = tx.objectStore('models');
-
-                            var rq = null;
-                            try {
-                                var rq = os.put(data, url);
-                            } catch (e) {
-                                cbPrint('loadRemote: failed to store "' + url + '" in the IndexedDB: \n' + e);
-                                cbCancel();
-                                return;
-                            }
+                            var rq = os.put(data, url);
 
                             rq.onsuccess = function (event) {
                                 cbPrint('loadRemote: "' + url + '" stored in the IndexedDB');
@@ -188,6 +180,7 @@ function loadRemote(url, dst, size_mb, cbProgress, cbReady, cbCancel, cbPrint) {
 
     rq.onabort = function (event) {
         cbPrint('loadRemote: failed to open IndexedDB: abort');
-        cbCancel();
+
     };
 }
+

examples/main/README.md

@@ -31,6 +31,7 @@ options:
   -osrt,     --output-srt        [false  ] output result in a srt file
   -owts,     --output-words      [false  ] output script for generating karaoke video
   -ocsv,     --output-csv        [false  ] output result in a CSV file
+  -oj,       --output-json       [false  ] output result in a JSON file
   -of FNAME, --output-file FNAME [       ] output file path (without file extension)
   -ps,       --print-special     [false  ] print special tokens
   -pc,       --print-colors      [false  ] print colors

examples/main/main.cpp

@@ -73,6 +73,7 @@ struct whisper_params {
     bool output_srt     = false;
     bool output_wts     = false;
     bool output_csv     = false;
+    bool output_jsn     = false;
     bool print_special  = false;
     bool print_colors   = false;
     bool print_progress = false;
@@ -80,6 +81,7 @@ struct whisper_params {
 
     std::string language = "en";
     std::string prompt;
+    std::string font_path = "/System/Library/Fonts/Supplemental/Courier New Bold.ttf";
     std::string model    = "models/ggml-base.en.bin";
 
     std::vector<std::string> fname_inp = {};
@@ -127,7 +129,9 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
         else if (arg == "-ovtt" || arg == "--output-vtt")     { params.output_vtt     = true; }
         else if (arg == "-osrt" || arg == "--output-srt")     { params.output_srt     = true; }
         else if (arg == "-owts" || arg == "--output-words")   { params.output_wts     = true; }
+        else if (arg == "-fp"   || arg == "--font-path")      { params.font_path      = argv[++i]; }
         else if (arg == "-ocsv" || arg == "--output-csv")     { params.output_csv     = true; }
+        else if (arg == "-oj"   || arg == "--output-json")    { params.output_jsn     = true; }
         else if (arg == "-of"   || arg == "--output-file")    { params.fname_out.emplace_back(argv[++i]); }
         else if (arg == "-ps"   || arg == "--print-special")  { params.print_special  = true; }
         else if (arg == "-pc"   || arg == "--print-colors")   { params.print_colors   = true; }
@@ -174,7 +178,9 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -ovtt,     --output-vtt        [%-7s] output result in a vtt file\n",                    params.output_vtt ? "true" : "false");
     fprintf(stderr, "  -osrt,     --output-srt        [%-7s] output result in a srt file\n",                    params.output_srt ? "true" : "false");
     fprintf(stderr, "  -owts,     --output-words      [%-7s] output script for generating karaoke video\n",     params.output_wts ? "true" : "false");
+    fprintf(stderr, "  -fp,       --font-path         [%-7s] path to a monospace font for karaoke video\n",     params.font_path.c_str());
     fprintf(stderr, "  -ocsv,     --output-csv        [%-7s] output result in a CSV file\n",                    params.output_csv ? "true" : "false");
+    fprintf(stderr, "  -oj,       --output-json       [%-7s] output result in a JSON file\n",                   params.output_jsn ? "true" : "false");
     fprintf(stderr, "  -of FNAME, --output-file FNAME [%-7s] output file path (without file extension)\n",      "");
     fprintf(stderr, "  -ps,       --print-special     [%-7s] print special tokens\n",                           params.print_special ? "true" : "false");
     fprintf(stderr, "  -pc,       --print-colors      [%-7s] print colors\n",                                   params.print_colors ? "true" : "false");
@@ -193,7 +199,7 @@ struct whisper_print_user_data {
     const std::vector<std::vector<float>> * pcmf32s;
 };
 
-void whisper_print_segment_callback(struct whisper_context * ctx, int n_new, void * user_data) {
+void whisper_print_segment_callback(struct whisper_context * ctx, struct whisper_state * /*state*/, int n_new, void * user_data) {
     const auto & params  = *((whisper_print_user_data *) user_data)->params;
     const auto & pcmf32s = *((whisper_print_user_data *) user_data)->pcmf32s;
 
@@ -352,28 +358,157 @@ bool output_csv(struct whisper_context * ctx, const char * fname) {
     fprintf(stderr, "%s: saving output to '%s'\n", __func__, fname);
 
     const int n_segments = whisper_full_n_segments(ctx);
+    fout << "start,end,text\n";
     for (int i = 0; i < n_segments; ++i) {
         const char * text = whisper_full_get_segment_text(ctx, i);
         const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
         const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
 
         //need to multiply times returned from whisper_full_get_segment_t{0,1}() by 10 to get milliseconds.
-        fout << 10 * t0 << ", " << 10 * t1 << ", \"" << text    << "\"\n";
+        fout << 10 * t0 << "," << 10 * t1 << ",\"" << text    << "\"\n";
     }
 
     return true;
 }
 
+bool output_json(struct whisper_context * ctx, const char * fname, const whisper_params & params) {
+    std::ofstream fout(fname);
+    int indent = 0;
+
+    auto doindent = [&]() {
+        for (int i = 0; i < indent; i++) fout << "\t";
+    };
+
+    auto start_arr = [&](const char *name) {
+        doindent();
+        fout << "\"" << name << "\": [\n";
+        indent++;
+    };
+
+    auto end_arr = [&](bool end = false) {
+        indent--;
+        doindent();
+        fout << (end ? "]\n" : "},\n");
+    };
+
+    auto start_obj = [&](const char *name = nullptr) {
+        doindent();
+        if (name) {
+            fout << "\"" << name << "\": {\n";
+        } else {
+            fout << "{\n";
+        }
+        indent++;
+    };
+
+    auto end_obj = [&](bool end = false) {
+        indent--;
+        doindent();
+        fout << (end ? "}\n" : "},\n");
+    };
+
+    auto start_value = [&](const char *name) {
+        doindent();
+        fout << "\"" << name << "\": ";
+    };
+
+    auto value_s = [&](const char *name, const char *val, bool end = false) {
+        start_value(name);
+        fout << "\"" << val << (end ? "\"\n" : "\",\n");
+    };
+
+    auto end_value = [&](bool end = false) {
+        fout << (end ? "\n" : ",\n");
+    };
+
+    auto value_i = [&](const char *name, const int64_t val, bool end = false) {
+        start_value(name);
+        fout << val;
+        end_value(end);
+    };
+
+    auto value_b = [&](const char *name, const bool val, bool end = false) {
+        start_value(name);
+        fout << (val ? "true" : "false");
+        end_value(end);
+    };
+
+    if (!fout.is_open()) {
+        fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname);
+        return false;
+    }
+
+    fprintf(stderr, "%s: saving output to '%s'\n", __func__, fname);
+    start_obj();
+        value_s("systeminfo", whisper_print_system_info());
+        start_obj("model");
+            value_s("type", whisper_model_type_readable(ctx));
+            value_b("multilingual", whisper_is_multilingual(ctx));
+            value_i("vocab", whisper_model_n_vocab(ctx));
+            start_obj("audio");
+                value_i("ctx", whisper_model_n_audio_ctx(ctx));
+                value_i("state", whisper_model_n_audio_state(ctx));
+                value_i("head", whisper_model_n_audio_head(ctx));
+                value_i("layer", whisper_model_n_audio_layer(ctx), true);
+            end_obj();
+            start_obj("text");
+                value_i("ctx", whisper_model_n_text_ctx(ctx));
+                value_i("state", whisper_model_n_text_state(ctx));
+                value_i("head", whisper_model_n_text_head(ctx));
+                value_i("leyer", whisper_model_n_text_layer(ctx), true);
+            end_obj();
+            value_i("mels", whisper_model_n_mels(ctx));
+            value_i("f16", whisper_model_f16(ctx), true);
+        end_obj();
+        start_obj("params");
+            value_s("model", params.model.c_str());
+            value_s("language", params.language.c_str());
+            value_b("translate", params.translate, true);
+        end_obj();
+        start_obj("result");
+            value_s("language", whisper_lang_str(whisper_full_lang_id(ctx)), true);
+        end_obj();
+        start_arr("transcription");
+
+            const int n_segments = whisper_full_n_segments(ctx);
+            for (int i = 0; i < n_segments; ++i) {
+                const char * text = whisper_full_get_segment_text(ctx, i);
+                const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
+                const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
+
+                start_obj();
+                    start_obj("timestanps");
+                        value_s("from", to_timestamp(t0, true).c_str());
+                        value_s("to", to_timestamp(t1, true).c_str(), true);
+                    end_obj();
+                    start_obj("offsets");
+                        value_i("from", t0 * 10);
+                        value_i("to", t1 * 10, true);
+                    end_obj();
+                    value_s("text", text, true);
+                end_obj(i == (n_segments - 1));
+            }
+
+        end_arr(true);
+    end_obj(true);
+    return true;
+}
+
 // karaoke video generation
 // outputs a bash script that uses ffmpeg to generate a video with the subtitles
 // TODO: font parameter adjustments
-bool output_wts(struct whisper_context * ctx, const char * fname, const char * fname_inp, const whisper_params & /*params*/, float t_sec) {
+bool output_wts(struct whisper_context * ctx, const char * fname, const char * fname_inp, const whisper_params & params, float t_sec) {
     std::ofstream fout(fname);
 
     fprintf(stderr, "%s: saving output to '%s'\n", __func__, fname);
 
-    // TODO: become parameter
-    static const char * font = "/System/Library/Fonts/Supplemental/Courier New Bold.ttf";
+    static const char * font = params.font_path.c_str();
+
+    std::ifstream fin(font);
+    if (!fin.is_open()) {
+        fprintf(stderr, "%s: font not found at '%s', please specify a monospace font with -fp\n", __func__, font);
+        return false;
+    }
 
     fout << "#!/bin/bash" << "\n";
     fout << "\n";
@@ -607,7 +742,7 @@ int main(int argc, char ** argv) {
             {
                 static bool is_aborted = false; // NOTE: this should be atomic to avoid data race
 
-                wparams.encoder_begin_callback = [](struct whisper_context * /*ctx*/, void * user_data) {
+                wparams.encoder_begin_callback = [](struct whisper_context * /*ctx*/, struct whisper_state * /*state*/, void * user_data) {
                     bool is_aborted = *(bool*)user_data;
                     return !is_aborted;
                 };
@@ -653,6 +788,12 @@ int main(int argc, char ** argv) {
                 const auto fname_csv = fname_out + ".csv";
                 output_csv(ctx, fname_csv.c_str());
             }
+
+            // output to JSON file
+            if (params.output_jsn) {
+                const auto fname_jsn = fname_out + ".json";
+                output_json(ctx, fname_jsn.c_str(), params);
+            }
         }
     }
 

examples/stream/stream.cpp

@@ -288,7 +288,6 @@ int main(int argc, char ** argv) {
             wparams.print_realtime   = false;
             wparams.print_timestamps = !params.no_timestamps;
             wparams.translate        = params.translate;
-            wparams.no_context       = true;
             wparams.single_segment   = !use_vad;
             wparams.max_tokens       = params.max_tokens;
             wparams.language         = params.language.c_str();

examples/talk/README.md

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

examples/whisper.android/README.md

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

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

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

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

@@ -41,10 +41,15 @@ class MainScreenViewModel(private val application: Application) : ViewModel() {
 
     init {
         viewModelScope.launch {
+            printSystemInfo()
             loadData()
         }
     }
 
+    private suspend fun printSystemInfo() {
+        printMessage(String.format("System Info: %s\n", WhisperContext.getSystemInfo()));
+    }
+
     private suspend fun loadData() {
         printMessage("Loading data...\n")
         try {
@@ -81,10 +86,29 @@ class MainScreenViewModel(private val application: Application) : ViewModel() {
         //whisperContext = WhisperContext.createContextFromFile(firstModel.absolutePath)
     }
 
+    fun benchmark() = viewModelScope.launch {
+        runBenchmark(6)
+    }
+
     fun transcribeSample() = viewModelScope.launch {
         transcribeAudio(getFirstSample())
     }
 
+    private suspend fun runBenchmark(nthreads: Int) {
+        if (!canTranscribe) {
+            return
+        }
+
+        canTranscribe = false
+
+        printMessage("Running benchmark. This will take minutes...\n")
+        whisperContext?.benchMemory(nthreads)?.let{ printMessage(it) }
+        printMessage("\n")
+        whisperContext?.benchGgmlMulMat(nthreads)?.let{ printMessage(it) }
+
+        canTranscribe = true
+    }
+
     private suspend fun getFirstSample(): File = withContext(Dispatchers.IO) {
         samplesPath.listFiles()!!.first()
     }
@@ -114,11 +138,14 @@ class MainScreenViewModel(private val application: Application) : ViewModel() {
         canTranscribe = false
 
         try {
-            printMessage("Reading wave samples...\n")
+            printMessage("Reading wave samples... ")
             val data = readAudioSamples(file)
+            printMessage("${data.size / (16000 / 1000)} ms\n")
             printMessage("Transcribing data...\n")
+            val start = System.currentTimeMillis()
             val text = whisperContext?.transcribeData(data)
-            printMessage("Done: $text\n")
+            val elapsed = System.currentTimeMillis() - start
+            printMessage("Done ($elapsed ms): $text\n")
         } catch (e: Exception) {
             Log.w(LOG_TAG, e)
             printMessage("${e.localizedMessage}\n")

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

@@ -27,6 +27,14 @@ class WhisperContext private constructor(private var ptr: Long) {
         }
     }
 
+    suspend fun benchMemory(nthreads: Int): String = withContext(scope.coroutineContext) {
+        return@withContext WhisperLib.benchMemcpy(nthreads)
+    }
+
+    suspend fun benchGgmlMulMat(nthreads: Int): String = withContext(scope.coroutineContext) {
+        return@withContext WhisperLib.benchGgmlMulMat(nthreads)
+    }
+
     suspend fun release() = withContext(scope.coroutineContext) {
         if (ptr != 0L) {
             WhisperLib.freeContext(ptr)
@@ -66,6 +74,10 @@ class WhisperContext private constructor(private var ptr: Long) {
             }
             return WhisperContext(ptr)
         }
+
+        fun getSystemInfo(): String {
+            return WhisperLib.getSystemInfo()
+        }
     }
 }
 
@@ -74,6 +86,7 @@ private class WhisperLib {
         init {
             Log.d(LOG_TAG, "Primary ABI: ${Build.SUPPORTED_ABIS[0]}")
             var loadVfpv4 = false
+            var loadV8fp16 = false
             if (isArmEabiV7a()) {
                 // armeabi-v7a needs runtime detection support
                 val cpuInfo = cpuInfo()
@@ -84,11 +97,24 @@ private class WhisperLib {
                         loadVfpv4 = true
                     }
                 }
+            } else if (isArmEabiV8a()) {
+                // ARMv8.2a needs runtime detection support
+                val cpuInfo = cpuInfo()
+                cpuInfo?.let {
+                    Log.d(LOG_TAG, "CPU info: $cpuInfo")
+                    if (cpuInfo.contains("fphp")) {
+                        Log.d(LOG_TAG, "CPU supports fp16 arithmetic")
+                        loadV8fp16 = true
+                    }
+                }
             }
 
             if (loadVfpv4) {
                 Log.d(LOG_TAG, "Loading libwhisper_vfpv4.so")
                 System.loadLibrary("whisper_vfpv4")
+            } else if (loadV8fp16) {
+                Log.d(LOG_TAG, "Loading libwhisper_v8fp16_va.so")
+                System.loadLibrary("whisper_v8fp16_va")
             } else {
                 Log.d(LOG_TAG, "Loading libwhisper.so")
                 System.loadLibrary("whisper")
@@ -103,6 +129,9 @@ private class WhisperLib {
         external fun fullTranscribe(contextPtr: Long, audioData: FloatArray)
         external fun getTextSegmentCount(contextPtr: Long): Int
         external fun getTextSegment(contextPtr: Long, index: Int): String
+        external fun getSystemInfo(): String
+        external fun benchMemcpy(nthread: Int): String
+        external fun benchGgmlMulMat(nthread: Int): String
     }
 }
 
@@ -110,6 +139,10 @@ private fun isArmEabiV7a(): Boolean {
     return Build.SUPPORTED_ABIS[0].equals("armeabi-v7a")
 }
 
+private fun isArmEabiV8a(): Boolean {
+    return Build.SUPPORTED_ABIS[0].equals("arm64-v8a")
+}
+
 private fun cpuInfo(): String? {
     return try {
         File("/proc/cpuinfo").inputStream().bufferedReader().use {

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

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

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

@@ -6,6 +6,7 @@
 #include <sys/sysinfo.h>
 #include <string.h>
 #include "whisper.h"
+#include "ggml.h"
 
 #define UNUSED(x) (void)(x)
 #define TAG "JNI"
@@ -213,4 +214,30 @@ Java_com_whispercppdemo_whisper_WhisperLib_00024Companion_getTextSegment(
     const char *text = whisper_full_get_segment_text(context, index);
     jstring string = (*env)->NewStringUTF(env, text);
     return string;
-}
+}
+
+JNIEXPORT jstring JNICALL
+Java_com_whispercppdemo_whisper_WhisperLib_00024Companion_getSystemInfo(
+        JNIEnv *env, jobject thiz
+) {
+    UNUSED(thiz);
+    const char *sysinfo = whisper_print_system_info();
+    jstring string = (*env)->NewStringUTF(env, sysinfo);
+    return string;
+}
+
+JNIEXPORT jstring JNICALL
+Java_com_whispercppdemo_whisper_WhisperLib_00024Companion_benchMemcpy(JNIEnv *env, jobject thiz,
+                                                                      jint n_threads) {
+    UNUSED(thiz);
+    const char *bench_ggml_memcpy = whisper_bench_memcpy_str(n_threads);
+    jstring string = (*env)->NewStringUTF(env, bench_ggml_memcpy);
+}
+
+JNIEXPORT jstring JNICALL
+Java_com_whispercppdemo_whisper_WhisperLib_00024Companion_benchGgmlMulMat(JNIEnv *env, jobject thiz,
+                                                                          jint n_threads) {
+    UNUSED(thiz);
+    const char *bench_ggml_mul_mat = whisper_bench_ggml_mul_mat_str(n_threads);
+    jstring string = (*env)->NewStringUTF(env, bench_ggml_mul_mat);
+}

examples/whisper.objc/README.md

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

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

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

examples/whisper.swiftui/README.md

@@ -7,8 +7,9 @@ To use:
 2. Add the model to "whisper.swiftui.demo/Resources/models" via Xcode.
 3. Select a sample audio file (for example, [jfk.wav](https://github.com/ggerganov/whisper.cpp/raw/master/samples/jfk.wav)).
 4. Add the model to "whisper.swiftui.demo/Resources/samples" via Xcode.
-5. Select the "release" build configuration under "Run", then deploy and run to your device.
+5. Select the "Release" [^2] build configuration under "Run", then deploy and run to your device.
 
 [^1]: I recommend the tiny, base or small models for running on an iOS device.
+[^2]: The `Release` build can boost performance of transcription. In this project, it also added `-O3 -DNDEBUG` to `Other C Flags`, but adding flags to app proj is not ideal in real world (applies to all C/C++ files), consider splitting xcodeproj in workspace in your own project.
 
 ![image](https://user-images.githubusercontent.com/1991296/212539216-0aef65e4-f882-480a-8358-0f816838fd52.png)

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

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

examples/whisper.wasm/CMakeLists.txt

@@ -31,9 +31,9 @@ endif()
 set_target_properties(${TARGET} PROPERTIES LINK_FLAGS " \
     --bind \
     -s USE_PTHREADS=1 \
-    -s PTHREAD_POOL_SIZE_STRICT=0 \
-    -s INITIAL_MEMORY=2000MB \
-    -s TOTAL_MEMORY=2000MB \
+    -s PTHREAD_POOL_SIZE=8 \
+    -s INITIAL_MEMORY=1500MB \
+    -s TOTAL_MEMORY=1500MB \
     -s FORCE_FILESYSTEM=1 \
     -s EXPORTED_RUNTIME_METHODS=\"['print', 'printErr', 'ccall', 'cwrap']\" \
     ${EXTRA_FLAGS} \

examples/whisper.wasm/emscripten.cpp

@@ -10,12 +10,6 @@ std::thread g_worker;
 
 std::vector<struct whisper_context *> g_contexts(4, nullptr);
 
-static inline int mpow2(int n) {
-    int p = 1;
-    while (p <= n) p *= 2;
-    return p/2;
-}
-
 EMSCRIPTEN_BINDINGS(whisper) {
     emscripten::function("init", emscripten::optional_override([](const std::string & path_model) {
         if (g_worker.joinable()) {
@@ -49,7 +43,7 @@ EMSCRIPTEN_BINDINGS(whisper) {
         }
     }));
 
-    emscripten::function("full_default", emscripten::optional_override([](size_t index, const emscripten::val & audio, const std::string & lang, int nthreads, bool translate) {
+    emscripten::function("full_default", emscripten::optional_override([](size_t index, const emscripten::val & audio, const std::string & lang, bool translate) {
         if (g_worker.joinable()) {
             g_worker.join();
         }
@@ -72,7 +66,7 @@ EMSCRIPTEN_BINDINGS(whisper) {
         params.print_special    = false;
         params.translate        = translate;
         params.language         = whisper_is_multilingual(g_contexts[index]) ? lang.c_str() : "en";
-        params.n_threads        = std::min(nthreads, std::min(16, mpow2(std::thread::hardware_concurrency())));
+        params.n_threads        = std::min(8, (int) std::thread::hardware_concurrency());
         params.offset_ms        = 0;
 
         std::vector<float> pcmf32;

examples/whisper.wasm/index-tmpl.html

@@ -40,34 +40,21 @@
 
             Note that the computation is quite heavy and may take a few seconds to complete.<br>
             The transcription results will be displayed in the text area below.<br><br>
-            <b>Important:</b>
-                <ul>
-                    <li>your browser must support WASM SIMD instructions for this to work</li>
-                    <li>quantized models are still in experimental stage (<a href="https://github.com/ggerganov/ggml/pull/27">more info</a>)</li>
-                    <li>Firefox cannot load files larger than 256 MB - use Chrome instead</li>
-                </ul>
+            <b>Important: your browser must support WASM SIMD instructions for this to work.</b>
 
-            <hr>
+            <br><br><hr>
 
             <div id="model">
-                Whisper models: <span id="model-whisper-status"></span><br><br>
+                Whisper model: <span id="model-whisper-status"></span>
                 <button id="fetch-whisper-tiny-en"  onclick="loadWhisper('tiny.en')">tiny.en (75 MB)</button>
                 <button id="fetch-whisper-tiny"     onclick="loadWhisper('tiny')">tiny (75 MB)</button>
                 <button id="fetch-whisper-base-en"  onclick="loadWhisper('base.en')">base.en (142 MB)</button>
                 <button id="fetch-whisper-base"     onclick="loadWhisper('base')">base (142 MB)</button>
                 <button id="fetch-whisper-small-en" onclick="loadWhisper('small.en')">small.en (466 MB)</button>
                 <button id="fetch-whisper-small"    onclick="loadWhisper('small')">small (466 MB)</button>
-                <input type="file" id="whisper-file" name="file" onchange="loadFile(event, 'whisper.bin')" />
-                <br><br>
-                Quantized models:<br><br>
-                <button id="fetch-whisper-base-en-q4_0"   onclick="loadWhisper('base-en-q4_0')">base.en (4bit, 49 MB)</button>
-                <button id="fetch-whisper-base-q4_0"      onclick="loadWhisper('base-q4_0')">base (4bit, 49 MB)</button>
-                <button id="fetch-whisper-small-en-q4_0"  onclick="loadWhisper('small-en-q4_0')">small.en (4bit, 152 MB)</button>
-                <button id="fetch-whisper-small-q4_0"     onclick="loadWhisper('small-q4_0')">small (4bit, 152 MB)</button><br>
-                <button id="fetch-whisper-medium-en-q4_0" onclick="loadWhisper('medium-en-q4_0')">medium.en (4bit, 469 MB)</button>
-                <button id="fetch-whisper-medium-q4_0"    onclick="loadWhisper('medium-q4_0')">medium (4bit, 469 MB)</button>
-                <button id="fetch-whisper-large-q4_0"     onclick="loadWhisper('large-q4_0')">large (4bit, 985 MB)</button>
                 <span id="fetch-whisper-progress"></span>
+
+                <input type="file" id="whisper-file" name="file" onchange="loadFile(event, 'whisper.bin')" />
             </div>
 
             <br>
@@ -174,12 +161,6 @@
                             <option value="yi">Yiddish</option>
                         </select>
                     </td>
-                    <!-- Slider to select number of threads between 1 and 16 -->
-                    <td>
-                        Threads:
-                        <input type="range" id="threads" name="threads" min="1" max="16" value="8" onchange="changeThreads(this.value)" />
-                        <span id="threads-value">8</span>
-                    </td>
                     <td>
                         <button onclick="onProcess(false);">Transcribe</button>
                     </td>
@@ -282,13 +263,11 @@
 
                 Module.FS_createDataFile("/", fname, buf, true, true);
 
-                //model_whisper = fname;
+                model_whisper = fname;
 
                 document.getElementById('model-whisper-status').innerHTML = 'loaded "' + model_whisper + '"!';
 
                 printTextarea('storeFS: stored model: ' + fname + ' size: ' + buf.length);
-
-                document.getElementById('model').innerHTML = 'Model fetched: ' + model_whisper;
             }
 
             function loadFile(event, fname) {
@@ -313,15 +292,6 @@
                 document.getElementById('fetch-whisper-tiny'    ).style.display = 'none';
                 document.getElementById('fetch-whisper-base'    ).style.display = 'none';
                 document.getElementById('fetch-whisper-small'   ).style.display = 'none';
-
-                document.getElementById('fetch-whisper-base-en-q4_0'  ).style.display = 'none';
-                document.getElementById('fetch-whisper-base-q4_0'     ).style.display = 'none';
-                document.getElementById('fetch-whisper-small-en-q4_0' ).style.display = 'none';
-                document.getElementById('fetch-whisper-small-q4_0'    ).style.display = 'none';
-                document.getElementById('fetch-whisper-medium-en-q4_0').style.display = 'none';
-                document.getElementById('fetch-whisper-medium-q4_0'   ).style.display = 'none';
-                document.getElementById('fetch-whisper-large-q4_0'    ).style.display = 'none';
-
                 document.getElementById('whisper-file'          ).style.display = 'none';
                 document.getElementById('model-whisper-status'  ).innerHTML = 'loaded model: ' + file.name;
             }
@@ -334,14 +304,6 @@
                     'base':     'https://whisper.ggerganov.com/ggml-model-whisper-base.bin',
                     'small.en': 'https://whisper.ggerganov.com/ggml-model-whisper-small.en.bin',
                     'small':    'https://whisper.ggerganov.com/ggml-model-whisper-small.bin',
-
-                    'base-en-q4_0':  'https://whisper.ggerganov.com/ggml-model-whisper-base.en-q4_0.bin',
-                    'base-q4_0':     'https://whisper.ggerganov.com/ggml-model-whisper-base-q4_0.bin',
-                    'small-en-q4_0': 'https://whisper.ggerganov.com/ggml-model-whisper-small.en-q4_0.bin',
-                    'small-q4_0':    'https://whisper.ggerganov.com/ggml-model-whisper-small-q4_0.bin',
-                    'medium-en-q4_0':'https://whisper.ggerganov.com/ggml-model-whisper-medium.en-q4_0.bin',
-                    'medium-q4_0':   'https://whisper.ggerganov.com/ggml-model-whisper-medium-q4_0.bin',
-                    'large-q4_0':    'https://whisper.ggerganov.com/ggml-model-whisper-large-q4_0.bin',
                 };
 
                 let sizes = {
@@ -351,14 +313,6 @@
                     'base':     142,
                     'small.en': 466,
                     'small':    466,
-
-                    'base-en-q4_0':   49,
-                    'base-q4_0':      49,
-                    'small-en-q4_0':  152,
-                    'small-q4_0':     152,
-                    'medium-en-q4_0': 469,
-                    'medium-q4_0':    469,
-                    'large-q4_0':     985,
                 };
 
                 let url     = urls[model];
@@ -373,15 +327,6 @@
                 document.getElementById('fetch-whisper-tiny'    ).style.display = 'none';
                 document.getElementById('fetch-whisper-base'    ).style.display = 'none';
                 document.getElementById('fetch-whisper-small'   ).style.display = 'none';
-
-                document.getElementById('fetch-whisper-base-en-q4_0'  ).style.display = 'none';
-                document.getElementById('fetch-whisper-base-q4_0'     ).style.display = 'none';
-                document.getElementById('fetch-whisper-small-en-q4_0' ).style.display = 'none';
-                document.getElementById('fetch-whisper-small-q4_0'    ).style.display = 'none';
-                document.getElementById('fetch-whisper-medium-en-q4_0').style.display = 'none';
-                document.getElementById('fetch-whisper-medium-q4_0'   ).style.display = 'none';
-                document.getElementById('fetch-whisper-large-q4_0'    ).style.display = 'none';
-
                 document.getElementById('whisper-file'          ).style.display = 'none';
                 document.getElementById('model-whisper-status'  ).innerHTML = 'loading model: ' + model;
 
@@ -392,22 +337,12 @@
 
                 cbCancel = function() {
                     var el;
-
                     el = document.getElementById('fetch-whisper-tiny-en' ); if (el) el.style.display = 'inline-block';
                     el = document.getElementById('fetch-whisper-base-en' ); if (el) el.style.display = 'inline-block';
                     el = document.getElementById('fetch-whisper-small-en'); if (el) el.style.display = 'inline-block';
                     el = document.getElementById('fetch-whisper-tiny'    ); if (el) el.style.display = 'inline-block';
                     el = document.getElementById('fetch-whisper-base'    ); if (el) el.style.display = 'inline-block';
                     el = document.getElementById('fetch-whisper-small'   ); if (el) el.style.display = 'inline-block';
-
-                    el = document.getElementById('fetch-whisper-base-en-q4_0'  ); if (el) el.style.display = 'inline-block';
-                    el = document.getElementById('fetch-whisper-base-q4_0'     ); if (el) el.style.display = 'inline-block';
-                    el = document.getElementById('fetch-whisper-small-en-q4_0' ); if (el) el.style.display = 'inline-block';
-                    el = document.getElementById('fetch-whisper-small-q4_0'    ); if (el) el.style.display = 'inline-block';
-                    el = document.getElementById('fetch-whisper-medium-en-q4_0'); if (el) el.style.display = 'inline-block';
-                    el = document.getElementById('fetch-whisper-medium-q4_0'   ); if (el) el.style.display = 'inline-block';
-                    el = document.getElementById('fetch-whisper-large-q4_0'    ); if (el) el.style.display = 'inline-block';
-
                     el = document.getElementById('whisper-file'          ); if (el) el.style.display = 'inline-block';
                     el = document.getElementById('model-whisper-status'  ); if (el) el.innerHTML = '';
                 };
@@ -419,8 +354,7 @@
             // audio file
             //
 
-            const kMaxAudio_s = 30*60;
-            const kMaxRecording_s = 2*60;
+            const kMaxAudio_s = 120;
             const kSampleRate = 16000;
 
             window.AudioContext = window.AudioContext || window.webkitAudioContext;
@@ -489,7 +423,7 @@
                 doRecording = false;
             }
 
-            // record up to kMaxRecording_s seconds of audio from the microphone
+            // record up to kMaxAudio_s seconds of audio from the microphone
             // check if doRecording is false every 1000 ms and stop recording if so
             // update progress information
             function startRecording() {
@@ -545,9 +479,9 @@
                                         printTextarea('js: audio recorded, size: ' + audio.length);
 
                                         // truncate to first 30 seconds
-                                        if (audio.length > kMaxRecording_s*kSampleRate) {
-                                            audio = audio.slice(0, kMaxRecording_s*kSampleRate);
-                                            printTextarea('js: truncated audio to first ' + kMaxRecording_s + ' seconds');
+                                        if (audio.length > kMaxAudio_s*kSampleRate) {
+                                            audio = audio.slice(0, kMaxAudio_s*kSampleRate);
+                                            printTextarea('js: truncated audio to first ' + kMaxAudio_s + ' seconds');
                                         }
                                         setAudio(audio);
                                     });
@@ -575,31 +509,24 @@
                         });
                     }
 
-                    document.getElementById('progress-bar').style.width = (100*(Date.now() - startTime)/1000/kMaxRecording_s) + '%';
-                    document.getElementById('progress-text').innerHTML = (100*(Date.now() - startTime)/1000/kMaxRecording_s).toFixed(0) + '%';
+                    document.getElementById('progress-bar').style.width = (100*(Date.now() - startTime)/1000/kMaxAudio_s) + '%';
+                    document.getElementById('progress-text').innerHTML = (100*(Date.now() - startTime)/1000/kMaxAudio_s).toFixed(0) + '%';
                 }, 1000);
 
                 printTextarea('js: recording ...');
 
                 setTimeout(function() {
                     if (doRecording) {
-                        printTextarea('js: recording stopped after ' + kMaxRecording_s + ' seconds');
+                        printTextarea('js: recording stopped after ' + kMaxAudio_s + ' seconds');
                         stopRecording();
                     }
-                }, kMaxRecording_s*1000);
+                }, kMaxAudio_s*1000);
             }
 
             //
             // transcribe
             //
 
-            var nthreads = 8;
-
-            function changeThreads(value) {
-                nthreads = value;
-                document.getElementById('threads-value').innerHTML = nthreads;
-            }
-
             function onProcess(translate) {
                 if (!instance) {
                     instance = Module.init('whisper.bin');
@@ -626,7 +553,7 @@
                     printTextarea('');
 
                     setTimeout(function() {
-                        var ret = Module.full_default(instance, audio, document.getElementById('language').value, nthreads, translate);
+                        var ret = Module.full_default(instance, audio, document.getElementById('language').value, translate);
                         console.log('js: full_default returned: ' + ret);
                         if (ret) {
                             printTextarea("js: whisper returned: " + ret);

extra/bench-wts.sh

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

ggml.h

@@ -198,8 +198,6 @@ struct ggml_object;
 struct ggml_context;
 
 enum ggml_type {
-    GGML_TYPE_Q4_0,
-    GGML_TYPE_Q4_1,
     GGML_TYPE_I8,
     GGML_TYPE_I16,
     GGML_TYPE_I32,
@@ -328,10 +326,7 @@ void ggml_print_objects(const struct ggml_context * ctx);
 int    ggml_nelements(const struct ggml_tensor * tensor);
 size_t ggml_nbytes   (const struct ggml_tensor * tensor);
 
-int    ggml_blck_size (enum ggml_type type);
-size_t ggml_type_size (enum ggml_type type); // size in bytes for all elements in a block
-float  ggml_type_sizef(enum ggml_type type); // ggml_type_size()/ggml_blck_size() as float
-
+size_t ggml_type_size   (enum ggml_type type);
 size_t ggml_element_size(const struct ggml_tensor * tensor);
 
 struct ggml_context * ggml_init(struct ggml_init_params params);

models/README.md

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

models/convert-h5-to-ggml.py

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

models/download-coreml-model.sh

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

models/download-ggml-model.cmd

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

models/download-ggml-model.sh

@@ -6,7 +6,7 @@
 #src="https://ggml.ggerganov.com"
 #pfx="ggml-model-whisper"
 
-src="https://huggingface.co/datasets/ggerganov/whisper.cpp"
+src="https://huggingface.co/ggerganov/whisper.cpp"
 pfx="resolve/main/ggml"
 
 # get the path of this script

whisper.h

@@ -66,6 +66,7 @@ extern "C" {
     //
 
     struct whisper_context;
+    struct whisper_state;
 
     typedef int whisper_token;
 
@@ -101,11 +102,20 @@ extern "C" {
     WHISPER_API struct whisper_context * whisper_init_from_buffer(void * buffer, size_t buffer_size);
     WHISPER_API struct whisper_context * whisper_init(struct whisper_model_loader * loader);
 
-    // Frees all memory allocated by the model.
-    WHISPER_API void whisper_free(struct whisper_context * ctx);
+    // These are the same as the above, but the internal state of the context is not allocated automatically
+    // It is the responsibility of the caller to allocate the state using whisper_init_state() (#523)
+    WHISPER_API struct whisper_context * whisper_init_from_file_no_state(const char * path_model);
+    WHISPER_API struct whisper_context * whisper_init_from_buffer_no_state(void * buffer, size_t buffer_size);
+    WHISPER_API struct whisper_context * whisper_init_no_state(struct whisper_model_loader * loader);
+
+    WHISPER_API struct whisper_state * whisper_init_state(struct whisper_context * ctx);
+
+    // Frees all allocated memory
+    WHISPER_API void whisper_free      (struct whisper_context * ctx);
+    WHISPER_API void whisper_free_state(struct whisper_state * state);
 
     // Convert RAW PCM audio to log mel spectrogram.
-    // The resulting spectrogram is stored inside the provided whisper context.
+    // The resulting spectrogram is stored inside the default state of the provided whisper context.
     // Returns 0 on success
     WHISPER_API int whisper_pcm_to_mel(
             struct whisper_context * ctx,
@@ -113,17 +123,30 @@ extern "C" {
                                int   n_samples,
                                int   n_threads);
 
-    // Convert RAW PCM audio to log mel spectrogram but applies a Phase Vocoder to speed up the audio x2. 
-    // The resulting spectrogram is stored inside the provided whisper context.
+    WHISPER_API int whisper_pcm_to_mel_with_state(
+            struct whisper_context * ctx,
+              struct whisper_state * state,
+                       const float * samples,
+                               int   n_samples,
+                               int   n_threads);
+
+    // Convert RAW PCM audio to log mel spectrogram but applies a Phase Vocoder to speed up the audio x2.
+    // The resulting spectrogram is stored inside the default state of the provided whisper context.
     // Returns 0 on success
     WHISPER_API int whisper_pcm_to_mel_phase_vocoder(
-        struct whisper_context* ctx,
-        const float* samples,
-        int   n_samples,
-        int   n_threads);
+        struct whisper_context * ctx,
+                   const float * samples,
+                           int   n_samples,
+                           int   n_threads);
 
+    WHISPER_API int whisper_pcm_to_mel_phase_vocoder_with_state(
+        struct whisper_context * ctx,
+          struct whisper_state * state,
+                   const float * samples,
+                           int   n_samples,
+                           int   n_threads);
 
-    // This can be used to set a custom log mel spectrogram inside the provided whisper context.
+    // This can be used to set a custom log mel spectrogram inside the default state of the provided whisper context.
     // Use this instead of whisper_pcm_to_mel() if you want to provide your own log mel spectrogram.
     // n_mel must be 80
     // Returns 0 on success
@@ -133,7 +156,14 @@ extern "C" {
                                int   n_len,
                                int   n_mel);
 
-    // Run the Whisper encoder on the log mel spectrogram stored inside the provided whisper context.
+    WHISPER_API int whisper_set_mel_with_state(
+            struct whisper_context * ctx,
+              struct whisper_state * state,
+                       const float * data,
+                               int   n_len,
+                               int   n_mel);
+
+    // Run the Whisper encoder on the log mel spectrogram stored inside the default state in the provided whisper context.
     // Make sure to call whisper_pcm_to_mel() or whisper_set_mel() first.
     // offset can be used to specify the offset of the first frame in the spectrogram.
     // Returns 0 on success
@@ -142,6 +172,12 @@ extern "C" {
                                int   offset,
                                int   n_threads);
 
+    WHISPER_API int whisper_encode_with_state(
+            struct whisper_context * ctx,
+              struct whisper_state * state,
+                               int   offset,
+                               int   n_threads);
+
     // Run the Whisper decoder to obtain the logits and probabilities for the next token.
     // Make sure to call whisper_encode() first.
     // tokens + n_tokens is the provided context for the decoder.
@@ -155,6 +191,14 @@ extern "C" {
                                int   n_past,
                                int   n_threads);
 
+    WHISPER_API int whisper_decode_with_state(
+            struct whisper_context * ctx,
+              struct whisper_state * state,
+               const whisper_token * tokens,
+                               int   n_tokens,
+                               int   n_past,
+                               int   n_threads);
+
     // Convert the provided text into tokens.
     // The tokens pointer must be large enough to hold the resulting tokens.
     // Returns the number of tokens on success, no more than n_max_tokens
@@ -190,20 +234,44 @@ extern "C" {
                                int   n_threads,
                              float * lang_probs);
 
-    WHISPER_API int whisper_n_len          (struct whisper_context * ctx); // mel length
-    WHISPER_API int whisper_n_vocab        (struct whisper_context * ctx);
-    WHISPER_API int whisper_n_text_ctx     (struct whisper_context * ctx);
-    WHISPER_API int whisper_n_audio_ctx    (struct whisper_context * ctx);
-    WHISPER_API int whisper_is_multilingual(struct whisper_context * ctx);
+    WHISPER_API int whisper_lang_auto_detect_with_state(
+            struct whisper_context * ctx,
+              struct whisper_state * state,
+                               int   offset_ms,
+                               int   n_threads,
+                             float * lang_probs);
+
+    WHISPER_API int whisper_n_len           (struct whisper_context * ctx); // mel length
+    WHISPER_API int whisper_n_len_from_state(struct whisper_state * state); // mel length
+    WHISPER_API int whisper_n_vocab         (struct whisper_context * ctx);
+    WHISPER_API int whisper_n_text_ctx      (struct whisper_context * ctx);
+    WHISPER_API int whisper_n_audio_ctx     (struct whisper_context * ctx);
+    WHISPER_API int whisper_is_multilingual (struct whisper_context * ctx);
+
+    WHISPER_API int whisper_model_n_vocab      (struct whisper_context * ctx);
+    WHISPER_API int whisper_model_n_audio_ctx  (struct whisper_context * ctx);
+    WHISPER_API int whisper_model_n_audio_state(struct whisper_context * ctx);
+    WHISPER_API int whisper_model_n_audio_head (struct whisper_context * ctx);
+    WHISPER_API int whisper_model_n_audio_layer(struct whisper_context * ctx);
+    WHISPER_API int whisper_model_n_text_ctx   (struct whisper_context * ctx);
+    WHISPER_API int whisper_model_n_text_state (struct whisper_context * ctx);
+    WHISPER_API int whisper_model_n_text_head  (struct whisper_context * ctx);
+    WHISPER_API int whisper_model_n_text_layer (struct whisper_context * ctx);
+    WHISPER_API int whisper_model_n_mels       (struct whisper_context * ctx);
+    WHISPER_API int whisper_model_f16          (struct whisper_context * ctx);
+    WHISPER_API int whisper_model_type         (struct whisper_context * ctx);
 
     // Token logits obtained from the last call to whisper_decode()
     // The logits for the last token are stored in the last row
     // Rows: n_tokens
     // Cols: n_vocab
-    WHISPER_API float * whisper_get_logits(struct whisper_context * ctx);
+    WHISPER_API float * whisper_get_logits           (struct whisper_context * ctx);
+    WHISPER_API float * whisper_get_logits_from_state(struct whisper_state * state);
 
     // Token Id -> String. Uses the vocabulary in the provided context
     WHISPER_API const char * whisper_token_to_str(struct whisper_context * ctx, whisper_token token);
+    WHISPER_API const char * whisper_model_type_readable(struct whisper_context * ctx);
+
 
     // Special tokens
     WHISPER_API whisper_token whisper_token_eot (struct whisper_context * ctx);
@@ -218,7 +286,7 @@ extern "C" {
     WHISPER_API whisper_token whisper_token_translate (void);
     WHISPER_API whisper_token whisper_token_transcribe(void);
 
-    // Performance information
+    // Performance information from the default state.
     WHISPER_API void whisper_print_timings(struct whisper_context * ctx);
     WHISPER_API void whisper_reset_timings(struct whisper_context * ctx);
 
@@ -236,18 +304,19 @@ extern "C" {
     // Text segment callback
     // Called on every newly generated text segment
     // Use the whisper_full_...() functions to obtain the text segments
-    typedef void (*whisper_new_segment_callback)(struct whisper_context * ctx, int n_new, void * user_data);
+    typedef void (*whisper_new_segment_callback)(struct whisper_context * ctx, struct whisper_state * state, int n_new, void * user_data);
 
     // Encoder begin callback
     // If not NULL, called before the encoder starts
     // If it returns false, the computation is aborted
-    typedef bool (*whisper_encoder_begin_callback)(struct whisper_context * ctx, void * user_data);
+    typedef bool (*whisper_encoder_begin_callback)(struct whisper_context * ctx, struct whisper_state * state, void * user_data);
 
     // Logits filter callback
     // Can be used to modify the logits before sampling
     // If not NULL, called after applying temperature to logits
     typedef void (*whisper_logits_filter_callback)(
             struct whisper_context * ctx,
+              struct whisper_state * state,
           const whisper_token_data * tokens,
                                int   n_tokens,
                              float * logits,
@@ -334,6 +403,7 @@ extern "C" {
     WHISPER_API struct whisper_full_params whisper_full_default_params(enum whisper_sampling_strategy strategy);
 
     // Run the entire model: PCM -> log mel spectrogram -> encoder -> decoder -> text
+    // Not thread safe for same context
     // Uses the specified decoding strategy to obtain the text.
     WHISPER_API int whisper_full(
                 struct whisper_context * ctx,
@@ -341,7 +411,16 @@ extern "C" {
                            const float * samples,
                                    int   n_samples);
 
-    // Split the input audio in chunks and process each chunk separately using whisper_full()
+    WHISPER_API int whisper_full_with_state(
+                struct whisper_context * ctx,
+                  struct whisper_state * state,
+            struct whisper_full_params   params,
+                           const float * samples,
+                                   int   n_samples);
+
+    // Split the input audio in chunks and process each chunk separately using whisper_full_with_state()
+    // Result is stored in the default state of the context
+    // Not thread safe if executed in parallel on the same context.
     // It seems this approach can offer some speedup in some cases.
     // However, the transcription accuracy can be worse at the beginning and end of each chunk.
     WHISPER_API int whisper_full_parallel(
@@ -351,40 +430,56 @@ extern "C" {
                                    int   n_samples,
                                    int   n_processors);
 
-    // Number of generated text segments.
+    // Number of generated text segments
     // A segment can be a few words, a sentence, or even a paragraph.
-    WHISPER_API int whisper_full_n_segments(struct whisper_context * ctx);
+    WHISPER_API int whisper_full_n_segments           (struct whisper_context * ctx);
+    WHISPER_API int whisper_full_n_segments_from_state(struct whisper_state * state);
 
-    // Language id associated with the current context
+    // Language id associated with the context's default state
     WHISPER_API int whisper_full_lang_id(struct whisper_context * ctx);
 
-    // Get the start and end time of the specified segment.
-    WHISPER_API int64_t whisper_full_get_segment_t0(struct whisper_context * ctx, int i_segment);
-    WHISPER_API int64_t whisper_full_get_segment_t1(struct whisper_context * ctx, int i_segment);
+    // Language id associated with the provided state
+    WHISPER_API int whisper_full_lang_id_from_state(struct whisper_state * state);
 
-    // Get the text of the specified segment.
-    WHISPER_API const char * whisper_full_get_segment_text(struct whisper_context * ctx, int i_segment);
+    // Get the start and end time of the specified segment
+    WHISPER_API int64_t whisper_full_get_segment_t0           (struct whisper_context * ctx, int i_segment);
+    WHISPER_API int64_t whisper_full_get_segment_t0_from_state(struct whisper_state * state, int i_segment);
 
-    // Get number of tokens in the specified segment.
-    WHISPER_API int whisper_full_n_tokens(struct whisper_context * ctx, int i_segment);
+    WHISPER_API int64_t whisper_full_get_segment_t1           (struct whisper_context * ctx, int i_segment);
+    WHISPER_API int64_t whisper_full_get_segment_t1_from_state(struct whisper_state * state, int i_segment);
 
-    // Get the token text of the specified token in the specified segment.
-    WHISPER_API const char * whisper_full_get_token_text(struct whisper_context * ctx, int i_segment, int i_token);
-    WHISPER_API whisper_token whisper_full_get_token_id (struct whisper_context * ctx, int i_segment, int i_token);
+    // Get the text of the specified segment
+    WHISPER_API const char * whisper_full_get_segment_text           (struct whisper_context * ctx, int i_segment);
+    WHISPER_API const char * whisper_full_get_segment_text_from_state(struct whisper_state * state, int i_segment);
 
-    // Get token data for the specified token in the specified segment.
+    // Get number of tokens in the specified segment
+    WHISPER_API int whisper_full_n_tokens           (struct whisper_context * ctx, int i_segment);
+    WHISPER_API int whisper_full_n_tokens_from_state(struct whisper_state * state, int i_segment);
+
+    // Get the token text of the specified token in the specified segment
+    WHISPER_API const char * whisper_full_get_token_text           (struct whisper_context * ctx, int i_segment, int i_token);
+    WHISPER_API const char * whisper_full_get_token_text_from_state(struct whisper_context * ctx, struct whisper_state * state, int i_segment, int i_token);
+
+    WHISPER_API whisper_token whisper_full_get_token_id           (struct whisper_context * ctx, int i_segment, int i_token);
+    WHISPER_API whisper_token whisper_full_get_token_id_from_state(struct whisper_state * state, int i_segment, int i_token);
+
+    // Get token data for the specified token in the specified segment
     // This contains probabilities, timestamps, etc.
-    WHISPER_API whisper_token_data whisper_full_get_token_data(struct whisper_context * ctx, int i_segment, int i_token);
+    WHISPER_API whisper_token_data whisper_full_get_token_data           (struct whisper_context * ctx, int i_segment, int i_token);
+    WHISPER_API whisper_token_data whisper_full_get_token_data_from_state(struct whisper_state * state, int i_segment, int i_token);
 
-    // Get the probability of the specified token in the specified segment.
-    WHISPER_API float whisper_full_get_token_p(struct whisper_context * ctx, int i_segment, int i_token);
+    // Get the probability of the specified token in the specified segment
+    WHISPER_API float whisper_full_get_token_p           (struct whisper_context * ctx, int i_segment, int i_token);
+    WHISPER_API float whisper_full_get_token_p_from_state(struct whisper_state * state, int i_segment, int i_token);
 
     ////////////////////////////////////////////////////////////////////////////
 
     // Temporary helpers needed for exposing ggml interface
 
     WHISPER_API int whisper_bench_memcpy(int n_threads);
+    WHISPER_API const char * whisper_bench_memcpy_str(int n_threads);
     WHISPER_API int whisper_bench_ggml_mul_mat(int n_threads);
+    WHISPER_API const char * whisper_bench_ggml_mul_mat_str(int n_threads);
 
 #ifdef __cplusplus
 }