minor

.gitignore

@@ -1,5 +1,7 @@
 *.o
 *.a
+*.mlmodel
+*.mlmodelc
 .cache/
 .vs/
 .vscode/
@@ -32,3 +34,5 @@ examples/whisper.objc/whisper.objc.xcodeproj/xcuserdata/
 examples/whisper.objc/whisper.objc.xcodeproj/project.xcworkspace/xcuserdata
 
 extra/bench-gg.txt
+
+*.mlmodel*

CMakeLists.txt

@@ -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()
 
@@ -183,6 +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
 #
@@ -202,6 +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

@@ -138,6 +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
@@ -190,11 +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
@@ -208,21 +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

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/CMakeLists.txt

@@ -63,4 +63,5 @@ else()
     add_subdirectory(command)
     add_subdirectory(bench)
     add_subdirectory(talk)
+    add_subdirectory(talk.llama)
 endif()

examples/talk.llama/.gitignore

@@ -0,0 +1,2 @@
+eleven-labs.py
+audio.mp3

examples/talk.llama/CMakeLists.txt

@@ -0,0 +1,12 @@
+if (WHISPER_SUPPORT_SDL2)
+    # talk.llama
+    set(TARGET talk-llama)
+
+    # TODO: this is temporary
+    #       need to export ggml symbols for MSVC, but too lazy ..
+    add_executable(${TARGET} talk-llama.cpp llama.cpp)
+
+    include(DefaultTargetOptions)
+
+    target_link_libraries(${TARGET} PRIVATE common common-sdl whisper ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
+endif ()

examples/talk.llama/README.md

@@ -0,0 +1,2 @@
+# talk.llama
+

examples/talk.llama/llama.h

@@ -0,0 +1,153 @@
+#ifndef LLAMA_H
+#define LLAMA_H
+
+#include <stddef.h>
+#include <stdint.h>
+#include <stdbool.h>
+
+#ifdef LLAMA_SHARED
+#    ifdef _WIN32
+#        ifdef LLAMA_BUILD
+#            define LLAMA_API __declspec(dllexport)
+#        else
+#            define LLAMA_API __declspec(dllimport)
+#        endif
+#    else
+#        define LLAMA_API __attribute__ ((visibility ("default")))
+#    endif
+#else
+#    define LLAMA_API
+#endif
+
+#define LLAMA_FILE_VERSION 1
+#define LLAMA_FILE_MAGIC 0x67676d66 // 'ggmf' in hex
+#define LLAMA_FILE_MAGIC_UNVERSIONED 0x67676d6c // pre-versioned files
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+    //
+    // C interface
+    //
+    // TODO: show sample usage
+    //
+
+    struct llama_context;
+
+    typedef int llama_token;
+
+    typedef struct llama_token_data {
+        llama_token id;  // token id
+
+        float p;     // probability of the token
+        float plog;  // log probability of the token
+
+    } llama_token_data;
+
+    typedef void (*llama_progress_callback)(double progress, void *ctx);
+
+    struct llama_context_params {
+        int n_ctx;   // text context
+        int n_parts; // -1 for default
+        int seed;    // RNG seed, 0 for random
+
+        bool f16_kv;     // use fp16 for KV cache
+        bool logits_all; // the llama_eval() call computes all logits, not just the last one
+        bool vocab_only; // only load the vocabulary, no weights
+        bool use_mlock;  // force system to keep model in RAM
+        bool embedding;  // embedding mode only
+
+        // called with a progress value between 0 and 1, pass NULL to disable
+        llama_progress_callback progress_callback;
+        // context pointer passed to the progress callback
+        void * progress_callback_user_data;
+    };
+
+    LLAMA_API struct llama_context_params llama_context_default_params();
+
+    // Various functions for loading a ggml llama model.
+    // Allocate (almost) all memory needed for the model.
+    // Return NULL on failure
+    LLAMA_API struct llama_context * llama_init_from_file(
+                             const char * path_model,
+            struct llama_context_params   params);
+
+    // Frees all allocated memory
+    LLAMA_API void llama_free(struct llama_context * ctx);
+
+    // TODO: not great API - very likely to change
+    // Returns 0 on success
+    LLAMA_API int llama_model_quantize(
+            const char * fname_inp,
+            const char * fname_out,
+                   int   itype,
+                   int   qk);
+
+    // Run the llama inference to obtain the logits and probabilities for the next token.
+    // tokens + n_tokens is the provided batch of new tokens to process
+    // n_past is the number of tokens to use from previous eval calls
+    // Returns 0 on success
+    LLAMA_API int llama_eval(
+            struct llama_context * ctx,
+               const llama_token * tokens,
+                             int   n_tokens,
+                             int   n_past,
+                             int   n_threads);
+
+    // Convert the provided text into tokens.
+    // The tokens pointer must be large enough to hold the resulting tokens.
+    // Returns the number of tokens on success, no more than n_max_tokens
+    // Returns a negative number on failure - the number of tokens that would have been returned
+    // TODO: not sure if correct
+    LLAMA_API int llama_tokenize(
+            struct llama_context * ctx,
+                      const char * text,
+                     llama_token * tokens,
+                             int   n_max_tokens,
+                            bool   add_bos);
+
+    LLAMA_API int llama_n_vocab(struct llama_context * ctx);
+    LLAMA_API int llama_n_ctx  (struct llama_context * ctx);
+    LLAMA_API int llama_n_embd (struct llama_context * ctx);
+
+    // Token logits obtained from the last call to llama_eval()
+    // The logits for the last token are stored in the last row
+    // Can be mutated in order to change the probabilities of the next token
+    // Rows: n_tokens
+    // Cols: n_vocab
+    LLAMA_API float * llama_get_logits(struct llama_context * ctx);
+
+    // Get the embeddings for the input
+    // shape: [n_embd] (1-dimensional)
+    LLAMA_API float * llama_get_embeddings(struct llama_context * ctx);
+
+    // Token Id -> String. Uses the vocabulary in the provided context
+    LLAMA_API const char * llama_token_to_str(struct llama_context * ctx, llama_token token);
+
+    // Special tokens
+    LLAMA_API llama_token llama_token_bos();
+    LLAMA_API llama_token llama_token_eos();
+
+    // TODO: improve the last_n_tokens interface ?
+    LLAMA_API llama_token llama_sample_top_p_top_k(
+       struct llama_context * ctx,
+          const llama_token * last_n_tokens_data,
+                        int   last_n_tokens_size,
+                        int   top_k,
+                     double   top_p,
+                     double   temp,
+                     double   repeat_penalty);
+
+    // Performance information
+    LLAMA_API void llama_print_timings(struct llama_context * ctx);
+    LLAMA_API void llama_reset_timings(struct llama_context * ctx);
+
+    // Print system information
+    LLAMA_API const char * llama_print_system_info(void);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif

examples/talk.llama/speak.sh

@@ -0,0 +1,20 @@
+#!/bin/bash
+
+# Usage:
+#  speak.sh <voice_id> <text-to-speak>
+
+# espeak
+# Mac OS: brew install espeak
+# Linux: apt-get install espeak
+#
+#espeak -v en-us+m$1 -s 225 -p 50 -a 200 -g 5 -k 5 "$2"
+
+# for Mac
+say "$2"
+
+# Eleven Labs
+#
+#wd=$(dirname $0)
+#script=$wd/eleven-labs.py
+#python3 $script $1 "$2" >/dev/null 2>&1
+#ffplay -autoexit -nodisp -loglevel quiet -hide_banner -i ./audio.mp3 >/dev/null 2>&1

examples/talk.llama/talk-llama.cpp

@@ -0,0 +1,511 @@
+// Talk with AI
+//
+
+#include "common.h"
+#include "common-sdl.h"
+#include "whisper.h"
+#include "llama.h"
+
+#include <cassert>
+#include <cstdio>
+#include <fstream>
+#include <regex>
+#include <string>
+#include <thread>
+#include <vector>
+#include <regex>
+
+std::vector<llama_token> llama_tokenize(struct llama_context * ctx, const std::string & text, bool add_bos) {
+    // initialize to prompt numer of chars, since n_tokens <= n_prompt_chars
+    std::vector<llama_token> res(text.size() + (int)add_bos);
+    int n = llama_tokenize(ctx, text.c_str(), res.data(), res.size(), add_bos);
+    assert(n >= 0);
+    res.resize(n);
+
+    return res;
+}
+
+// command-line parameters
+struct whisper_params {
+    int32_t n_threads  = std::min(4, (int32_t) std::thread::hardware_concurrency());
+    int32_t voice_ms   = 10000;
+    int32_t capture_id = -1;
+    int32_t max_tokens = 32;
+    int32_t audio_ctx  = 0;
+
+    float vad_thold    = 0.6f;
+    float freq_thold   = 100.0f;
+
+    bool speed_up      = false;
+    bool translate     = false;
+    bool print_special = false;
+    bool print_energy  = false;
+    bool no_timestamps = true;
+
+    std::string person      = "Santa";
+    std::string language    = "en";
+    std::string model_wsp   = "models/ggml-base.en.bin";
+    std::string model_llama = "models/ggml-llama-7B.bin";
+    std::string speak       = "./examples/talk/speak.sh";
+    std::string fname_out;
+};
+
+void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
+
+bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
+    for (int i = 1; i < argc; i++) {
+        std::string arg = argv[i];
+
+        if (arg == "-h" || arg == "--help") {
+            whisper_print_usage(argc, argv, params);
+            exit(0);
+        }
+        else if (arg == "-t"   || arg == "--threads")       { params.n_threads     = std::stoi(argv[++i]); }
+        else if (arg == "-vms" || arg == "--voice-ms")      { params.voice_ms      = std::stoi(argv[++i]); }
+        else if (arg == "-c"   || arg == "--capture")       { params.capture_id    = std::stoi(argv[++i]); }
+        else if (arg == "-mt"  || arg == "--max-tokens")    { params.max_tokens    = std::stoi(argv[++i]); }
+        else if (arg == "-ac"  || arg == "--audio-ctx")     { params.audio_ctx     = std::stoi(argv[++i]); }
+        else if (arg == "-vth" || arg == "--vad-thold")     { params.vad_thold     = std::stof(argv[++i]); }
+        else if (arg == "-fth" || arg == "--freq-thold")    { params.freq_thold    = std::stof(argv[++i]); }
+        else if (arg == "-su"  || arg == "--speed-up")      { params.speed_up      = true; }
+        else if (arg == "-tr"  || arg == "--translate")     { params.translate     = true; }
+        else if (arg == "-ps"  || arg == "--print-special") { params.print_special = true; }
+        else if (arg == "-pe"  || arg == "--print-energy")  { params.print_energy  = true; }
+        else if (arg == "-p"   || arg == "--person")        { params.person        = argv[++i]; }
+        else if (arg == "-l"   || arg == "--language")      { params.language      = argv[++i]; }
+        else if (arg == "-mw"  || arg == "--model-whisper") { params.model_wsp     = argv[++i]; }
+        else if (arg == "-ml"  || arg == "--model-llama")   { params.model_llama   = argv[++i]; }
+        else if (arg == "-s"   || arg == "--speak")         { params.speak         = argv[++i]; }
+        else if (arg == "-f"   || arg == "--file")          { params.fname_out     = argv[++i]; }
+        else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            whisper_print_usage(argc, argv, params);
+            exit(0);
+        }
+    }
+
+    return true;
+}
+
+void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & params) {
+    fprintf(stderr, "\n");
+    fprintf(stderr, "usage: %s [options]\n", argv[0]);
+    fprintf(stderr, "\n");
+    fprintf(stderr, "options:\n");
+    fprintf(stderr, "  -h,       --help          [default] show this help message and exit\n");
+    fprintf(stderr, "  -t N,     --threads N     [%-7d] number of threads to use during computation\n", params.n_threads);
+    fprintf(stderr, "  -vms N,   --voice-ms N    [%-7d] voice duration in milliseconds\n",              params.voice_ms);
+    fprintf(stderr, "  -c ID,    --capture ID    [%-7d] capture device ID\n",                           params.capture_id);
+    fprintf(stderr, "  -mt N,    --max-tokens N  [%-7d] maximum number of tokens per audio chunk\n",    params.max_tokens);
+    fprintf(stderr, "  -ac N,    --audio-ctx N   [%-7d] audio context size (0 - all)\n",                params.audio_ctx);
+    fprintf(stderr, "  -vth N,   --vad-thold N   [%-7.2f] voice activity detection threshold\n",        params.vad_thold);
+    fprintf(stderr, "  -fth N,   --freq-thold N  [%-7.2f] high-pass frequency cutoff\n",                params.freq_thold);
+    fprintf(stderr, "  -su,      --speed-up      [%-7s] speed up audio by x2 (reduced accuracy)\n",     params.speed_up ? "true" : "false");
+    fprintf(stderr, "  -tr,      --translate     [%-7s] translate from source language to english\n",   params.translate ? "true" : "false");
+    fprintf(stderr, "  -ps,      --print-special [%-7s] print special tokens\n",                        params.print_special ? "true" : "false");
+    fprintf(stderr, "  -pe,      --print-energy  [%-7s] print sound energy (for debugging)\n",          params.print_energy ? "true" : "false");
+    fprintf(stderr, "  -p NAME,  --person NAME   [%-7s] person name (for prompt selection)\n",          params.person.c_str());
+    fprintf(stderr, "  -l LANG,  --language LANG [%-7s] spoken language\n",                             params.language.c_str());
+    fprintf(stderr, "  -mw FILE, --model-whisper [%-7s] whisper model file\n",                          params.model_wsp.c_str());
+    fprintf(stderr, "  -mg FILE, --model-llama   [%-7s] llama model file\n",                            params.model_llama.c_str());
+    fprintf(stderr, "  -s FILE,  --speak TEXT    [%-7s] command for TTS\n",                             params.speak.c_str());
+    fprintf(stderr, "  -f FNAME, --file FNAME    [%-7s] text output file name\n",                       params.fname_out.c_str());
+    fprintf(stderr, "\n");
+}
+
+std::string transcribe(whisper_context * ctx, const whisper_params & params, const std::vector<float> & pcmf32, float & prob, int64_t & t_ms) {
+    const auto t_start = std::chrono::high_resolution_clock::now();
+
+    prob = 0.0f;
+    t_ms = 0;
+
+    whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
+
+    wparams.print_progress   = false;
+    wparams.print_special    = params.print_special;
+    wparams.print_realtime   = false;
+    wparams.print_timestamps = !params.no_timestamps;
+    wparams.translate        = params.translate;
+    wparams.no_context       = true;
+    wparams.single_segment   = true;
+    wparams.max_tokens       = params.max_tokens;
+    wparams.language         = params.language.c_str();
+    wparams.n_threads        = params.n_threads;
+
+    wparams.audio_ctx        = params.audio_ctx;
+    wparams.speed_up         = params.speed_up;
+
+    if (whisper_full(ctx, wparams, pcmf32.data(), pcmf32.size()) != 0) {
+        return "";
+    }
+
+    int prob_n = 0;
+    std::string result;
+
+    const int n_segments = whisper_full_n_segments(ctx);
+    for (int i = 0; i < n_segments; ++i) {
+        const char * text = whisper_full_get_segment_text(ctx, i);
+
+        result += text;
+
+        const int n_tokens = whisper_full_n_tokens(ctx, i);
+        for (int j = 0; j < n_tokens; ++j) {
+            const auto token = whisper_full_get_token_data(ctx, i, j);
+
+            prob += token.p;
+            ++prob_n;
+        }
+    }
+
+    if (prob_n > 0) {
+        prob /= prob_n;
+    }
+
+    const auto t_end = std::chrono::high_resolution_clock::now();
+    t_ms = std::chrono::duration_cast<std::chrono::milliseconds>(t_end - t_start).count();
+
+    return result;
+}
+
+// need to have leading ' '
+//const std::string k_prompt = R"( Transcript of a dialog, where {1} interacts with an Assistant named Bob. Bob is helpful, kind, honest, good at writing, and never fails to answer {1}'s requests immediately and with precision.
+//
+//{0}: Hello, Bob.
+//{1}: Hello {0}. How may I help you today?
+//{0}:)";
+
+const std::string k_prompt = R"( Text transcript of a never ending dialog, where {0} interacts with an AI assistant named {1}.
+{1} is helpful, kind, honest, friendly, good at writing and never fails to answer {0}’s requests immediately and with details and precision.
+There are no annotations like (30 seconds passed...) or (to himself), just what {0} and {1} say aloud to each other.
+The transcript only includes text, it does not include markup like HTML and Markdown.
+{1} answers responds with short and concise answers.
+
+{0}{4} Hello, {1}!
+{1}{4} Hello {0}! How may I help you today?
+{0}{4} What time is it?
+{1}{4} It is {2} o'clock.
+{0}{4} What year is it?
+{1}{4} We are in {3}.
+{0}{4} What is a cat?
+{1}{4} A cat is a domestic species of small carnivorous mammal. It is the only domesticated species in the family Felidae.
+{0}{4} Name a color.
+{1}{4} Blue
+{0}{4})";
+
+int main(int argc, char ** argv) {
+    whisper_params params;
+
+    if (whisper_params_parse(argc, argv, params) == false) {
+        return 1;
+    }
+
+    if (whisper_lang_id(params.language.c_str()) == -1) {
+        fprintf(stderr, "error: unknown language '%s'\n", params.language.c_str());
+        whisper_print_usage(argc, argv, params);
+        exit(0);
+    }
+
+    // whisper init
+
+    struct whisper_context * ctx_wsp = whisper_init_from_file(params.model_wsp.c_str());
+
+    // llama init
+
+    auto lparams = llama_context_default_params();
+
+    lparams.n_ctx      = 512;
+    lparams.n_parts    = 2; // TODO fix
+    lparams.seed       = 1; // TODO fix
+    lparams.f16_kv     = true;
+
+    struct llama_context * ctx_llama = llama_init_from_file(params.model_llama.c_str(), lparams);
+
+    // print some info about the processing
+    {
+        fprintf(stderr, "\n");
+        if (!whisper_is_multilingual(ctx_wsp)) {
+            if (params.language != "en" || params.translate) {
+                params.language = "en";
+                params.translate = false;
+                fprintf(stderr, "%s: WARNING: model is not multilingual, ignoring language and translation options\n", __func__);
+            }
+        }
+        fprintf(stderr, "%s: processing, %d threads, lang = %s, task = %s, timestamps = %d ...\n",
+                __func__,
+                params.n_threads,
+                params.language.c_str(),
+                params.translate ? "translate" : "transcribe",
+                params.no_timestamps ? 0 : 1);
+
+        fprintf(stderr, "\n");
+    }
+
+
+    // init audio
+
+    audio_async audio(30*1000);
+    if (!audio.init(params.capture_id, WHISPER_SAMPLE_RATE)) {
+        fprintf(stderr, "%s: audio.init() failed!\n", __func__);
+        return 1;
+    }
+
+    audio.resume();
+
+    int n_iter = 0;
+
+    bool is_running  = true;
+    bool force_speak = false;
+
+    float prob0 = 0.0f;
+
+    const std::string chat_symb = ":";
+    const std::string bot_name = "LLAMA";
+
+    std::vector<float> pcmf32_cur;
+    std::vector<float> pcmf32_prompt;
+
+    std::string prompt_org = k_prompt;
+    prompt_org = ::replace(prompt_org, "{0}", params.person);
+    prompt_org = ::replace(prompt_org, "{1}", bot_name);
+
+    {
+        // get time string
+        std::string time_str;
+        {
+            time_t t = time(0);
+            struct tm * now = localtime(&t);
+            char buf[128];
+            strftime(buf, sizeof(buf), "%H:%M", now);
+            time_str = buf;
+        }
+        prompt_org = ::replace(prompt_org, "{2}", time_str);
+    }
+
+    {
+        // get year string
+        std::string year_str;
+        {
+            time_t t = time(0);
+            struct tm * now = localtime(&t);
+            char buf[128];
+            strftime(buf, sizeof(buf), "%Y", now);
+            year_str = buf;
+        }
+        prompt_org = ::replace(prompt_org, "{3}", year_str);
+    }
+
+    prompt_org = ::replace(prompt_org, "{4}", chat_symb);
+
+    auto embd_inp = ::llama_tokenize(ctx_llama, prompt_org, true);
+
+    const int n_ctx = llama_n_ctx(ctx_llama);
+
+    printf("\n");
+    printf("%s : initializing - please wait ...\n", __func__);
+
+    if (llama_eval(ctx_llama, embd_inp.data(), embd_inp.size(), 0, params.n_threads)) {
+        fprintf(stderr, "%s : failed to eval\n", __func__);
+        return 1;
+    }
+
+    //fprintf(stdout, "\n");
+    //fprintf(stdout, "%s", prompt_org.c_str());
+    //fflush(stdout);
+
+    printf("%s : done! start speaking in the microphone\n", __func__);
+    printf("\n");
+    printf("%s%s", params.person.c_str(), chat_symb.c_str());
+    fflush(stdout);
+
+    audio.clear();
+
+    const int n_keep = embd_inp.size();
+    const int voice_id = 2;
+
+    int n_past = n_keep;
+    int n_prev = 64; // TODO arg
+
+    std::vector<llama_token> embd;
+
+    std::vector<std::string> antiprompts = {
+        params.person + chat_symb,
+    };
+
+    // main loop
+    while (is_running) {
+        // handle Ctrl + C
+        is_running = sdl_poll_events();
+
+        if (!is_running) {
+            break;
+        }
+
+        // delay
+        std::this_thread::sleep_for(std::chrono::milliseconds(100));
+
+        int64_t t_ms = 0;
+
+        {
+            audio.get(2000, pcmf32_cur);
+
+            if (::vad_simple(pcmf32_cur, WHISPER_SAMPLE_RATE, 1250, params.vad_thold, params.freq_thold, params.print_energy) || force_speak) {
+                //fprintf(stdout, "%s: Speech detected! Processing ...\n", __func__);
+
+                audio.get(params.voice_ms, pcmf32_cur);
+
+                std::string text_heard;
+
+                if (!force_speak) {
+                    text_heard = ::trim(::transcribe(ctx_wsp, params, pcmf32_cur, prob0, t_ms));
+                }
+
+                // remove text between brackets using regex
+                {
+                    std::regex re("\\[.*?\\]");
+                    text_heard = std::regex_replace(text_heard, re, "");
+                }
+
+                // remove text between brackets using regex
+                {
+                    std::regex re("\\(.*?\\)");
+                    text_heard = std::regex_replace(text_heard, re, "");
+                }
+
+                // remove all characters, except for letters, numbers, punctuation and ':', '\'', '-', ' '
+                text_heard = std::regex_replace(text_heard, std::regex("[^a-zA-Z0-9\\.,\\?!\\s\\:\\'\\-]"), "");
+
+                // take first line
+                text_heard = text_heard.substr(0, text_heard.find_first_of('\n'));
+
+                // remove leading and trailing whitespace
+                text_heard = std::regex_replace(text_heard, std::regex("^\\s+"), "");
+                text_heard = std::regex_replace(text_heard, std::regex("\\s+$"), "");
+
+                const std::vector<llama_token> tokens = llama_tokenize(ctx_llama, text_heard.c_str(), false);
+
+                if (text_heard.empty() || tokens.empty() || force_speak) {
+                    //fprintf(stdout, "%s: Heard nothing, skipping ...\n", __func__);
+                    audio.clear();
+
+                    continue;
+                }
+
+                force_speak = false;
+
+                text_heard.insert(0, 1, ' ');
+                text_heard += "\n" + bot_name + chat_symb;
+                fprintf(stdout, "%s%s%s", "\033[1m", text_heard.c_str(), "\033[0m");
+                fflush(stdout);
+
+                embd = ::llama_tokenize(ctx_llama, text_heard, false);
+
+                // text inference
+                bool done = false;
+                std::string text_to_speak;
+                while (true) {
+                    // predict
+                    if (embd.size() > 0) {
+                        if (n_past + (int) embd.size() > n_ctx) {
+                            n_past = n_keep;
+
+                            // insert n_left/2 tokens at the start of embd from last_n_tokens
+                            embd.insert(embd.begin(), embd_inp.begin() + embd_inp.size() - n_prev, embd_inp.end());
+
+                            //printf("\n---\n");
+                            //printf("resetting: '");
+                            //for (int i = 0; i < (int) embd.size(); i++) {
+                            //    printf("%s", llama_token_to_str(ctx_llama, embd[i]));
+                            //}
+                            //printf("'\n");
+                            //printf("\n---\n");
+                        }
+
+                        if (llama_eval(ctx_llama, embd.data(), embd.size(), n_past, params.n_threads)) {
+                            fprintf(stderr, "%s : failed to eval\n", __func__);
+                            return 1;
+                        }
+                    }
+
+                    //printf("n_iter = %d, n_past = %d, n_ctx = %d, n_keep = %d, n_prev = %d, embd.size() = %d\n", n_iter, n_past, n_ctx, n_keep, n_prev, (int) embd.size());
+
+                    embd_inp.insert(embd_inp.end(), embd.begin(), embd.end());
+                    n_past += embd.size();
+                    embd.clear();
+
+                    if (done) break;
+
+                    {
+                        // out of user input, sample next token
+                        const float top_k          = 5;
+                        const float top_p          = 0.80f;
+                        const float temp           = 0.30f;
+                        const float repeat_penalty = 1.1764f;
+
+                        const int repeat_last_n    = 256;
+
+                        llama_token id = 0;
+
+                        {
+                            //auto logits = llama_get_logits(ctx_llama);
+                            //logits[llama_token_eos()] = 0;
+
+                            id = llama_sample_top_p_top_k(ctx_llama,
+                                    embd_inp.data() + std::max(0, n_past - repeat_last_n),
+                                    repeat_last_n, top_k, top_p, temp, repeat_penalty);
+                        }
+
+                        if (id != llama_token_eos()) {
+                            // add it to the context
+                            embd.push_back(id);
+
+                            text_to_speak += llama_token_to_str(ctx_llama, id);
+
+                            printf("%s", llama_token_to_str(ctx_llama, id));
+                        } else {
+                            // TODO
+                            printf("EOS TOKEN - SHOULD NOT HAPPEN\n");
+                            exit(0);
+                        }
+                    }
+
+                    {
+                        std::string last_output;
+                        for (int i = embd_inp.size() - 16; i < (int) embd_inp.size(); i++) {
+                            last_output += llama_token_to_str(ctx_llama, embd_inp[i]);
+                        }
+                        last_output += llama_token_to_str(ctx_llama, embd[0]);
+
+                        for (std::string & antiprompt : antiprompts) {
+                            if (last_output.find(antiprompt.c_str(), last_output.length() - antiprompt.length(), antiprompt.length()) != std::string::npos) {
+                                done = true;
+                                text_to_speak = ::replace(text_to_speak, antiprompt, "");
+                                fflush(stdout);
+                                break;
+                            }
+                        }
+                    }
+
+                    is_running = sdl_poll_events();
+
+                    if (!is_running) {
+                        break;
+                    }
+                }
+
+                text_to_speak = ::replace(text_to_speak, "\"", "");
+                system((params.speak + " " + std::to_string(voice_id) + " \"" + text_to_speak + "\"").c_str());
+
+                audio.clear();
+
+                ++n_iter;
+            }
+        }
+    }
+
+    audio.pause();
+
+    whisper_print_timings(ctx_wsp);
+    whisper_free(ctx_wsp);
+
+    return 0;
+}

ggml.h

@@ -198,6 +198,8 @@ 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,
@@ -226,7 +228,9 @@ enum ggml_op {
     GGML_OP_STEP,
     GGML_OP_RELU,
     GGML_OP_GELU,
+    GGML_OP_SILU,
     GGML_OP_NORM, // normalize
+    GGML_OP_RMS_NORM,
 
     GGML_OP_MUL_MAT,
 
@@ -326,7 +330,10 @@ 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);
 
-size_t ggml_type_size   (enum ggml_type type);
+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_element_size(const struct ggml_tensor * tensor);
 
 struct ggml_context * ggml_init(struct ggml_init_params params);
@@ -336,6 +343,9 @@ size_t ggml_used_mem(const struct ggml_context * ctx);
 
 size_t ggml_set_scratch(struct ggml_context * ctx, struct ggml_scratch scratch);
 
+bool ggml_mlock_supported(void);
+bool ggml_mlock(struct ggml_context * ctx, char ** err_p);
+
 struct ggml_tensor * ggml_new_tensor(
         struct ggml_context * ctx,
         enum   ggml_type type,
@@ -466,12 +476,20 @@ struct ggml_tensor * ggml_gelu(
         struct ggml_context * ctx,
         struct ggml_tensor  * a);
 
+struct ggml_tensor * ggml_silu(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a);
+
 // normalize along rows
 // TODO: eps is hardcoded to 1e-5 for now
 struct ggml_tensor * ggml_norm(
         struct ggml_context * ctx,
         struct ggml_tensor  * a);
 
+struct ggml_tensor * ggml_rms_norm(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a);
+
 // A: m rows, n columns
 // B: p rows, n columns (i.e. we transpose it internally)
 // result is m columns, p rows
@@ -726,6 +744,13 @@ enum ggml_opt_result ggml_opt(
         struct ggml_opt_params params,
         struct ggml_tensor * f);
 
+//
+// quantization
+//
+
+size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int qk, int64_t * hist);
+size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int qk, int64_t * hist);
+
 //
 // system info
 //

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"

whisper.cpp

@@ -1,5 +1,8 @@
 #define WHISPER_BUILD
 #include "whisper.h"
+#if WHISPER_USE_COREML
+#include "coreml/whisper-encoder.h"
+#endif
 
 #include "ggml.h"
 
@@ -586,6 +589,10 @@ struct whisper_state {
 
     int lang_id = 0; // english by default
 
+#ifdef WHISPER_USE_COREML
+    whisper_coreml_context * ctx_coreml;
+#endif
+
     // [EXPERIMENTAL] token-level timestamps data
     int64_t t_beg = 0;
     int64_t t_last = 0;
@@ -636,6 +643,8 @@ struct whisper_context {
     whisper_model model;
     whisper_vocab vocab;
     whisper_state * state = nullptr;
+
+    std::string path_model; // populated by whisper_init_from_file()
 };
 
 template<typename T>
@@ -1366,6 +1375,7 @@ static bool whisper_encode_internal(
         }
     }
 
+#ifndef WHISPER_USE_COREML
     struct ggml_tensor * cur;
 
     // convolution + gelu
@@ -1597,7 +1607,7 @@ static bool whisper_encode_internal(
                         ggml_repeat(ctx0, layer.mlp_ln_w, cur),
                         cur),
                     ggml_repeat(ctx0, layer.mlp_ln_b, cur));
-    }
+            }
 
 #ifdef WHISPER_USE_FLASH_FF
             wstate.use_buf(ctx0, 0);
@@ -1637,7 +1647,7 @@ static bool whisper_encode_internal(
                 ggml_repeat(ctx0, layer.mlp_1_b, cur),
                 cur);
 #endif
-}
+        }
 
         wstate.use_buf(ctx0, 3);
 
@@ -1674,6 +1684,13 @@ static bool whisper_encode_internal(
 
         //ggml_graph_print(&gf);
     }
+#else
+    wstate.use_buf(ctx0, -1);
+
+    struct ggml_tensor * cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_state, n_ctx);
+
+    whisper_coreml_encode(wstate.ctx_coreml, (float *) mel->data, (float *) cur->data);
+#endif
 
     // cur
     //{
@@ -1841,8 +1858,6 @@ static bool whisper_decode_internal(
 
         // self-attention
         {
-            wstate.use_buf(ctx0, 1);
-
             struct ggml_tensor * Qcur = ggml_mul_mat(ctx0,
                     layer.attn_q_w,
                     cur);
@@ -1904,8 +1919,6 @@ static bool whisper_decode_internal(
             // K * Q
             struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
 
-            wstate.use_buf(ctx0, 0);
-
             //struct ggml_tensor * KQ_scaled =
             //    ggml_scale(ctx0,
             //            KQ,
@@ -1914,20 +1927,16 @@ static bool whisper_decode_internal(
 
             struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ, n_past);
 
-            wstate.use_buf(ctx0, 1);
-
             struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked);
 
-            wstate.use_buf(ctx0, 0);
-
             struct ggml_tensor * V_trans =
-                ggml_permute(ctx0,
-                        ggml_reshape_3d(ctx0,
-                            ggml_view_1d(ctx0, kv_self.v, (n_past + N)*n_state, il*n_ctx*ggml_element_size(kv_self.v)*n_state),
-                            n_state/n_head, n_head, n_past + N),
-                        1, 2, 0, 3);
-
-            wstate.use_buf(ctx0, 1);
+                ggml_cpy(ctx0,
+                        ggml_permute(ctx0,
+                            ggml_reshape_3d(ctx0,
+                                ggml_view_1d(ctx0, kv_self.v, (n_past + N)*n_state, il*n_ctx*ggml_element_size(kv_self.v)*n_state),
+                                n_state/n_head, n_head, n_past + N),
+                            1, 2, 0, 3),
+                        ggml_new_tensor_3d(ctx0, kv_self.v->type, n_past + N, n_state/n_head, n_head));
 
             struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_trans, KQ_soft_max);
 
@@ -1964,8 +1973,6 @@ static bool whisper_decode_internal(
 
             cur = ggml_norm(ctx0, inpCA); // note: we use inpCA here
 
-            wstate.use_buf(ctx0, 1);
-
             // cur = ln_0_w*cur + ln_0_b
             cur = ggml_add(ctx0,
                     ggml_mul(ctx0,
@@ -1976,8 +1983,6 @@ static bool whisper_decode_internal(
 
         // cross-attention
         {
-            wstate.use_buf(ctx0, 0);
-
             struct ggml_tensor * Qcur = ggml_mul_mat(ctx0,
                     layer.cross_attn_q_w,
                     cur);
@@ -2001,12 +2006,13 @@ static bool whisper_decode_internal(
                         ggml_view_1d(ctx0, wstate.kv_cross.v, M*n_state, il*M*ggml_element_size(wstate.kv_cross.v)*n_state),
                         n_state/n_head, n_head, M);
 
-            struct ggml_tensor * V_trans = ggml_permute(ctx0, Vcross, 1, 2, 0, 3);
+            struct ggml_tensor * V_trans =
+                ggml_cpy(ctx0,
+                        ggml_permute(ctx0, Vcross, 1, 2, 0, 3),
+                        ggml_new_tensor_3d(ctx0, Vcross->type, M, n_state/n_head, n_head));
 
             // ------
 
-            wstate.use_buf(ctx0, 1);
-
             struct ggml_tensor * Q =
                 ggml_permute(ctx0,
                         ggml_cpy(ctx0,
@@ -2016,8 +2022,6 @@ static bool whisper_decode_internal(
 
             struct ggml_tensor * K = ggml_permute(ctx0, Kcross, 0, 2, 1, 3);
 
-            wstate.use_buf(ctx0, 0);
-
             // K * Q
             struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
 
@@ -2030,16 +2034,10 @@ static bool whisper_decode_internal(
             // no masking for cross-attention
             //struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled, n_past);
 
-            wstate.use_buf(ctx0, 1);
-
             struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ);
 
-            wstate.use_buf(ctx0, 0);
-
             struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_trans, KQ_soft_max);
 
-            wstate.use_buf(ctx0, 1);
-
             struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
 
             // cur = KQV_merged.contiguous().view(n_state, N)
@@ -2477,12 +2475,25 @@ static std::vector<whisper_vocab::id> tokenize(const whisper_vocab & vocab, cons
 // interface implementation
 //
 
+#ifdef WHISPER_USE_COREML
+// replace .bin with .mlmodelc
+static std::string whisper_get_coreml_path(std::string path_bin) {
+    auto pos = path_bin.rfind('.');
+    if (pos != std::string::npos) {
+        path_bin = path_bin.substr(0, pos);
+    }
+
+    path_bin += ".mlmodelc";
+
+    return path_bin;
+}
+#endif
+
 struct whisper_state * whisper_init_state(whisper_context * ctx) {
     whisper_state * state = new whisper_state;
 
     const size_t scale = ctx->model.hparams.f16 ? 1 : 2;
 
-
     if (!kv_cache_init(ctx->model.hparams, scale * MEM_REQ_KV_SELF.at(ctx->model.type), state->decoders[0].kv_self, ctx->wtype, ctx->model.hparams.n_text_ctx)) {
         fprintf(stderr, "%s: kv_cache_init() failed for self-attention cache\n", __func__);
         return nullptr;
@@ -2503,7 +2514,6 @@ struct whisper_state * whisper_init_state(whisper_context * ctx) {
         fprintf(stderr, "%s: kv cross size = %7.2f MB\n", __func__, memory_size / 1024.0 / 1024.0);
     }
 
-
     state->logits.reserve(ctx->vocab.n_vocab * ctx->model.hparams.n_text_ctx);
 
     state->logits_id.reserve(ctx->model.hparams.n_vocab);
@@ -2523,6 +2533,21 @@ struct whisper_state * whisper_init_state(whisper_context * ctx) {
 
     state->rng = std::mt19937(0);
 
+#ifdef WHISPER_USE_COREML
+    const auto path_coreml = whisper_get_coreml_path(ctx->path_model);
+
+    fprintf(stderr, "%s: loading Core ML model from '%s'\n", __func__, path_coreml.c_str());
+    fprintf(stderr, "%s: first run on a device may take a while ...\n", __func__);
+
+    state->ctx_coreml = whisper_coreml_init(path_coreml.c_str());
+    if (!state->ctx_coreml) {
+        fprintf(stderr, "%s: failed to load Core ML model from '%s'\n", __func__, path_coreml.c_str());
+        return nullptr;
+    }
+
+    fprintf(stderr, "%s: Core ML model loaded\n", __func__);
+#endif
+
     return state;
 }
 
@@ -2538,6 +2563,7 @@ struct whisper_context * whisper_init_from_file_no_state(const char * path_model
     }
 
     loader.context = &fin;
+
     loader.read = [](void * ctx, void * output, size_t read_size) {
         std::ifstream * fin = (std::ifstream*)ctx;
         fin->read((char *)output, read_size);
@@ -2554,7 +2580,13 @@ struct whisper_context * whisper_init_from_file_no_state(const char * path_model
         fin->close();
     };
 
-    return whisper_init_no_state(&loader);
+    auto ctx = whisper_init_no_state(&loader);
+
+    if (ctx) {
+        ctx->path_model = path_model;
+    }
+
+    return ctx;
 }
 
 struct whisper_context * whisper_init_from_buffer_no_state(void * buffer, size_t buffer_size) {
@@ -2679,6 +2711,10 @@ void whisper_free(struct whisper_context * ctx) {
 
         whisper_free_state(ctx->state);
 
+#ifdef WHISPER_USE_COREML
+        whisper_coreml_free(ctx->state->ctx_coreml);
+        ctx->state->ctx_coreml = nullptr;
+#endif
         delete ctx;
     }
 }