whispercpp 1.3.1 → 1.3.2

data/ext/{include → sources/include}/whisper.h

@@ -189,6 +189,15 @@ extern "C" {
         uint32_t             value; // Unicode code point or rule ID
     } whisper_grammar_element;
 
+    typedef struct whisper_vad_params {
+        float threshold;               // Probability threshold to consider as speech.
+        int   min_speech_duration_ms;  // Min duration for a valid speech segment.
+        int   min_silence_duration_ms; // Min silence duration to consider speech as ended.
+        float max_speech_duration_s;   // Max duration of a speech segment before forcing a new segment.
+        int   speech_pad_ms;           // Padding added before and after speech segments.
+        float samples_overlap;         // Overlap in seconds when copying audio samples from speech segment.
+    } whisper_vad_params;
+
     // Various functions for loading a ggml whisper model.
     // Allocate (almost) all memory needed for the model.
     // Return NULL on failure
@@ -522,8 +531,8 @@ extern "C" {
         bool detect_language;
 
         // common decoding parameters:
-        bool suppress_blank;    // ref: https://github.com/openai/whisper/blob/f82bc59f5ea234d4b97fb2860842ed38519f7e65/whisper/decoding.py#L89
-        bool suppress_non_speech_tokens; // ref: https://github.com/openai/whisper/blob/7858aa9c08d98f75575035ecd6481f462d66ca27/whisper/tokenizer.py#L224-L253
+        bool suppress_blank; // ref: https://github.com/openai/whisper/blob/f82bc59f5ea234d4b97fb2860842ed38519f7e65/whisper/decoding.py#L89
+        bool suppress_nst;   // non-speech tokens, ref: https://github.com/openai/whisper/blob/7858aa9c08d98f75575035ecd6481f462d66ca27/whisper/tokenizer.py#L224-L253
 
         float temperature;      // initial decoding temperature, ref: https://ai.stackexchange.com/a/32478
         float max_initial_ts;   // ref: https://github.com/openai/whisper/blob/f82bc59f5ea234d4b97fb2860842ed38519f7e65/whisper/decoding.py#L97
@@ -570,11 +579,18 @@ extern "C" {
         size_t                           n_grammar_rules;
         size_t                           i_start_rule;
         float                            grammar_penalty;
+
+        // Voice Activity Detection (VAD) params
+        bool         vad;                         // Enable VAD
+        const char * vad_model_path;              // Path to VAD model
+
+        whisper_vad_params vad_params;
     };
 
     // NOTE: this function allocates memory, and it is the responsibility of the caller to free the pointer - see whisper_free_context_params & whisper_free_params()
     WHISPER_API struct whisper_context_params * whisper_context_default_params_by_ref(void);
     WHISPER_API struct whisper_context_params   whisper_context_default_params       (void);
+
     WHISPER_API struct whisper_full_params * whisper_full_default_params_by_ref(enum whisper_sampling_strategy strategy);
     WHISPER_API struct whisper_full_params   whisper_full_default_params       (enum whisper_sampling_strategy strategy);
 
@@ -652,6 +668,53 @@ extern "C" {
     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);
 
+    //
+    // Voice Activity Detection (VAD)
+    //
+
+    struct whisper_vad_context;
+
+    WHISPER_API struct whisper_vad_params whisper_vad_default_params(void);
+
+    struct whisper_vad_context_params {
+        int   n_threads;  // The number of threads to use for processing.
+        bool  use_gpu;
+        int   gpu_device; // CUDA device
+    };
+
+    WHISPER_API struct whisper_vad_context_params whisper_vad_default_context_params(void);
+
+    WHISPER_API struct whisper_vad_context * whisper_vad_init_from_file_with_params(const char * path_model,              struct whisper_vad_context_params params);
+    WHISPER_API struct whisper_vad_context * whisper_vad_init_with_params          (struct whisper_model_loader * loader, struct whisper_vad_context_params params);
+
+    WHISPER_API bool whisper_vad_detect_speech(
+            struct whisper_vad_context * vctx,
+                           const float * samples,
+                                   int   n_samples);
+
+    WHISPER_API int     whisper_vad_n_probs(struct whisper_vad_context * vctx);
+    WHISPER_API float * whisper_vad_probs  (struct whisper_vad_context * vctx);
+
+    struct whisper_vad_segments;
+
+    WHISPER_API struct whisper_vad_segments * whisper_vad_segments_from_probs(
+            struct whisper_vad_context * vctx,
+            struct whisper_vad_params    params);
+
+    WHISPER_API struct whisper_vad_segments * whisper_vad_segments_from_samples(
+            struct whisper_vad_context * vctx,
+            struct whisper_vad_params    params,
+                           const float * samples,
+                                   int   n_samples);
+
+    WHISPER_API int whisper_vad_segments_n_segments(struct whisper_vad_segments * segments);
+
+    WHISPER_API float whisper_vad_segments_get_segment_t0(struct whisper_vad_segments * segments, int i_segment);
+    WHISPER_API float whisper_vad_segments_get_segment_t1(struct whisper_vad_segments * segments, int i_segment);
+
+    WHISPER_API void whisper_vad_free_segments(struct whisper_vad_segments * segments);
+    WHISPER_API void whisper_vad_free         (struct whisper_vad_context  * ctx);
+
     ////////////////////////////////////////////////////////////////////////////
 
     // Temporary helpers needed for exposing ggml interface
@@ -665,6 +728,9 @@ extern "C" {
 
     WHISPER_API void whisper_log_set(ggml_log_callback log_callback, void * user_data);
 
+    // Get the no_speech probability for the specified segment
+    WHISPER_API float whisper_full_get_segment_no_speech_prob           (struct whisper_context * ctx, int i_segment);
+    WHISPER_API float whisper_full_get_segment_no_speech_prob_from_state(struct whisper_state * state, int i_segment);
 #ifdef __cplusplus
 }
 #endif

data/ext/sources/src/CMakeLists.txt

@@ -0,0 +1,143 @@
+# TODO: should not use this
+if (WIN32)
+    add_compile_definitions(_CRT_SECURE_NO_WARNINGS)
+
+    if (BUILD_SHARED_LIBS)
+        set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON)
+    endif()
+endif()
+
+
+if (WHISPER_ALL_WARNINGS)
+    if (NOT MSVC)
+        list(APPEND WARNING_FLAGS -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function)
+        list(APPEND C_FLAGS       -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmissing-prototypes
+                                  -Werror=implicit-int -Werror=implicit-function-declaration)
+        list(APPEND CXX_FLAGS     -Wmissing-declarations -Wmissing-noreturn)
+
+        list(APPEND C_FLAGS   ${WARNING_FLAGS})
+        list(APPEND CXX_FLAGS ${WARNING_FLAGS})
+
+        add_compile_options("$<$<COMPILE_LANGUAGE:C>:${C_FLAGS}>"
+                            "$<$<COMPILE_LANGUAGE:CXX>:${CXX_FLAGS}>")
+    else()
+        # todo : msvc
+        set(C_FLAGS   "")
+        set(CXX_FLAGS "")
+    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(FATAL_ERROR "CoreML framework not found")
+    endif()
+
+    if (WHISPER_COREML_ALLOW_FALLBACK)
+        set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DWHISPER_COREML_ALLOW_FALLBACK)
+    endif()
+endif()
+
+if (WHISPER_OPENVINO)
+    find_package(OpenVINO REQUIRED COMPONENTS Runtime)
+endif()
+
+#
+# libraries
+#
+
+# whisper.coreml
+
+if (WHISPER_COREML)
+    set(TARGET whisper.coreml)
+
+    add_library(${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"
+        XCODE_ATTRIBUTE_CLANG_ENABLE_OBJC_ARC YES
+        )
+    set_target_properties(${TARGET} PROPERTIES FOLDER "libs")
+endif()
+
+if (WHISPER_OPENVINO)
+    set(TARGET whisper.openvino)
+
+    add_library(${TARGET} OBJECT
+        openvino/whisper-openvino-encoder.h
+        openvino/whisper-openvino-encoder.cpp
+        )
+
+    target_include_directories(${TARGET} PUBLIC
+        .
+        )
+
+    set_property(TARGET ${TARGET} PROPERTY POSITION_INDEPENDENT_CODE ON)
+    set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DWHISPER_USE_OPENVINO)
+
+    target_link_libraries(${TARGET} PRIVATE ggml openvino::runtime)
+    set_target_properties(${TARGET} PROPERTIES FOLDER "libs")
+endif()
+
+# whisper
+
+add_library(whisper
+            ../include/whisper.h
+            whisper-arch.h
+            whisper.cpp
+            )
+
+# Set the version numbers
+set_target_properties(whisper PROPERTIES
+    VERSION ${PROJECT_VERSION}
+    SOVERSION ${SOVERSION}
+)
+
+target_include_directories(whisper PUBLIC . ../include)
+target_compile_features   (whisper PUBLIC cxx_std_11) # don't bump
+
+if (CMAKE_CXX_BYTE_ORDER STREQUAL "BIG_ENDIAN")
+    set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DWHISPER_BIG_ENDIAN)
+endif()
+
+if (WHISPER_EXTRA_FLAGS)
+    target_compile_options(whisper PRIVATE ${WHISPER_EXTRA_FLAGS})
+endif()
+
+target_link_libraries(whisper PUBLIC ggml)
+
+if (WHISPER_COREML)
+    target_link_libraries(whisper PRIVATE whisper.coreml)
+endif()
+
+if (WHISPER_OPENVINO)
+    target_link_libraries(whisper PRIVATE whisper.openvino)
+endif()
+
+if (WHISPER_MKL)
+    target_link_libraries(whisper PRIVATE MKL::MKL)
+endif()
+
+if (BUILD_SHARED_LIBS)
+    set_target_properties(whisper PROPERTIES POSITION_INDEPENDENT_CODE ON)
+    target_compile_definitions(whisper PRIVATE WHISPER_SHARED WHISPER_BUILD)
+endif()

data/ext/{src → sources/src}/coreml/whisper-decoder-impl.h

@@ -11,36 +11,33 @@
 
 NS_ASSUME_NONNULL_BEGIN
 
-
 /// Model Prediction Input Type
-API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((visibility("hidden")))
+API_AVAILABLE(macos(10.15), ios(13.0), watchos(6.0), tvos(13.0)) __attribute__((visibility("hidden")))
 @interface whisper_decoder_implInput : NSObject<MLFeatureProvider>
 
-/// token_data as 1 by 1 matrix of 32-bit integers
+/// token_data as 1 by 1 matrix of floats
 @property (readwrite, nonatomic, strong) MLMultiArray * token_data;
 
-/// audio_data as 1 × 384 × 1 × 1500 4-dimensional array of floats
+/// audio_data as 1 × 1500 × 384 3-dimensional array of floats
 @property (readwrite, nonatomic, strong) MLMultiArray * audio_data;
 - (instancetype)init NS_UNAVAILABLE;
 - (instancetype)initWithToken_data:(MLMultiArray *)token_data audio_data:(MLMultiArray *)audio_data NS_DESIGNATED_INITIALIZER;
 
 @end
 
-
 /// Model Prediction Output Type
-API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((visibility("hidden")))
+API_AVAILABLE(macos(10.15), ios(13.0), watchos(6.0), tvos(13.0)) __attribute__((visibility("hidden")))
 @interface whisper_decoder_implOutput : NSObject<MLFeatureProvider>
 
-/// var_1346 as multidimensional array of floats
-@property (readwrite, nonatomic, strong) MLMultiArray * var_1346;
+/// cast_76 as multidimensional array of floats
+@property (readwrite, nonatomic, strong) MLMultiArray * cast_76;
 - (instancetype)init NS_UNAVAILABLE;
-- (instancetype)initWithVar_1346:(MLMultiArray *)var_1346 NS_DESIGNATED_INITIALIZER;
+- (instancetype)initWithCast_76:(MLMultiArray *)cast_76 NS_DESIGNATED_INITIALIZER;
 
 @end
 
-
 /// Class for model loading and prediction
-API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((visibility("hidden")))
+API_AVAILABLE(macos(10.15), ios(13.0), watchos(6.0), tvos(13.0)) __attribute__((visibility("hidden")))
 @interface whisper_decoder_impl : NSObject
 @property (readonly, nonatomic, nullable) MLModel * model;
 
@@ -94,7 +91,7 @@ API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((v
     @param configuration The model configuration
     @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid whisper_decoder_impl instance or NSError object.
 */
-+ (void)loadWithConfiguration:(MLModelConfiguration *)configuration completionHandler:(void (^)(whisper_decoder_impl * _Nullable model, NSError * _Nullable error))handler;
++ (void)loadWithConfiguration:(MLModelConfiguration *)configuration completionHandler:(void (^)(whisper_decoder_impl * _Nullable model, NSError * _Nullable error))handler API_AVAILABLE(macos(11.0), ios(14.0), watchos(7.0), tvos(14.0)) __attribute__((visibility("hidden")));
 
 /**
     Construct whisper_decoder_impl instance asynchronously with URL of .mlmodelc directory and optional configuration.
@@ -105,7 +102,7 @@ API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((v
     @param configuration The model configuration
     @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid whisper_decoder_impl instance or NSError object.
 */
-+ (void)loadContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration completionHandler:(void (^)(whisper_decoder_impl * _Nullable model, NSError * _Nullable error))handler;
++ (void)loadContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration completionHandler:(void (^)(whisper_decoder_impl * _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
@@ -124,10 +121,25 @@ API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((v
 */
 - (nullable whisper_decoder_implOutput *)predictionFromFeatures:(whisper_decoder_implInput *)input options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error;
 
+/**
+    Make an asynchronous prediction using the standard interface
+    @param input an instance of whisper_decoder_implInput to predict from
+    @param completionHandler a block that will be called upon completion of the prediction. error will be nil if no error occurred.
+*/
+- (void)predictionFromFeatures:(whisper_decoder_implInput *)input completionHandler:(void (^)(whisper_decoder_implOutput * _Nullable output, NSError * _Nullable error))completionHandler API_AVAILABLE(macos(14.0), ios(17.0), watchos(10.0), tvos(17.0)) __attribute__((visibility("hidden")));
+
+/**
+    Make an asynchronous prediction using the standard interface
+    @param input an instance of whisper_decoder_implInput to predict from
+    @param options prediction options
+    @param completionHandler a block that will be called upon completion of the prediction. error will be nil if no error occurred.
+*/
+- (void)predictionFromFeatures:(whisper_decoder_implInput *)input options:(MLPredictionOptions *)options completionHandler:(void (^)(whisper_decoder_implOutput * _Nullable output, NSError * _Nullable error))completionHandler API_AVAILABLE(macos(14.0), ios(17.0), watchos(10.0), tvos(17.0)) __attribute__((visibility("hidden")));
+
 /**
     Make a prediction using the convenience interface
-    @param token_data as 1 by 1 matrix of 32-bit integers:
-    @param audio_data as 1 × 384 × 1 × 1500 4-dimensional array of floats:
+    @param token_data 1 by 1 matrix of floats
+    @param audio_data 1 × 1500 × 384 3-dimensional array of floats
     @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
     @return the prediction as whisper_decoder_implOutput
 */

data/ext/{src → sources/src}/coreml/whisper-decoder-impl.m

@@ -39,21 +39,21 @@
 
 @implementation whisper_decoder_implOutput
 
-- (instancetype)initWithVar_1346:(MLMultiArray *)var_1346 {
+- (instancetype)initWithCast_76:(MLMultiArray *)cast_76 {
     self = [super init];
     if (self) {
-        _var_1346 = var_1346;
+        _cast_76 = cast_76;
     }
     return self;
 }
 
 - (NSSet<NSString *> *)featureNames {
-    return [NSSet setWithArray:@[@"var_1346"]];
+    return [NSSet setWithArray:@[@"cast_76"]];
 }
 
 - (nullable MLFeatureValue *)featureValueForName:(NSString *)featureName {
-    if ([featureName isEqualToString:@"var_1346"]) {
-        return [MLFeatureValue featureValueWithMultiArray:self.var_1346];
+    if ([featureName isEqualToString:@"cast_76"]) {
+        return [MLFeatureValue featureValueWithMultiArray:self.cast_76];
     }
     return nil;
 }
@@ -80,10 +80,13 @@
     Such application may want to use `-[MLModel initWithContentsOfURL:configuration:error:]` and `+URLOfModelInThisBundle` to create a MLModel object to pass-in.
 */
 - (instancetype)initWithMLModel:(MLModel *)model {
+    if (model == nil) {
+        return nil;
+    }
     self = [super init];
-    if (!self) { return nil; }
-    _model = model;
-    if (_model == nil) { return nil; }
+    if (self != nil) {
+        _model = model;
+    }
     return self;
 }
 
@@ -177,7 +180,29 @@
 - (nullable whisper_decoder_implOutput *)predictionFromFeatures:(whisper_decoder_implInput *)input options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error {
     id<MLFeatureProvider> outFeatures = [self.model predictionFromFeatures:input options:options error:error];
     if (!outFeatures) { return nil; }
-    return [[whisper_decoder_implOutput alloc] initWithVar_1346:(MLMultiArray *)[outFeatures featureValueForName:@"var_1346"].multiArrayValue];
+    return [[whisper_decoder_implOutput alloc] initWithCast_76:(MLMultiArray *)[outFeatures featureValueForName:@"cast_76"].multiArrayValue];
+}
+
+- (void)predictionFromFeatures:(whisper_decoder_implInput *)input completionHandler:(void (^)(whisper_decoder_implOutput * _Nullable output, NSError * _Nullable error))completionHandler {
+    [self.model predictionFromFeatures:input completionHandler:^(id<MLFeatureProvider> prediction, NSError *predictionError) {
+        if (prediction != nil) {
+            whisper_decoder_implOutput *output = [[whisper_decoder_implOutput alloc] initWithCast_76:(MLMultiArray *)[prediction featureValueForName:@"cast_76"].multiArrayValue];
+            completionHandler(output, predictionError);
+        } else {
+            completionHandler(nil, predictionError);
+        }
+    }];
+}
+
+- (void)predictionFromFeatures:(whisper_decoder_implInput *)input options:(MLPredictionOptions *)options completionHandler:(void (^)(whisper_decoder_implOutput * _Nullable output, NSError * _Nullable error))completionHandler {
+    [self.model predictionFromFeatures:input options:options completionHandler:^(id<MLFeatureProvider> prediction, NSError *predictionError) {
+        if (prediction != nil) {
+            whisper_decoder_implOutput *output = [[whisper_decoder_implOutput alloc] initWithCast_76:(MLMultiArray *)[prediction featureValueForName:@"cast_76"].multiArrayValue];
+            completionHandler(output, predictionError);
+        } else {
+            completionHandler(nil, predictionError);
+        }
+    }];
 }
 
 - (nullable whisper_decoder_implOutput *)predictionFromToken_data:(MLMultiArray *)token_data audio_data:(MLMultiArray *)audio_data error:(NSError * _Nullable __autoreleasing * _Nullable)error {
@@ -192,7 +217,7 @@
     NSMutableArray<whisper_decoder_implOutput*> *results = [NSMutableArray arrayWithCapacity:(NSUInteger)outBatch.count];
     for (NSInteger i = 0; i < outBatch.count; i++) {
         id<MLFeatureProvider> resultProvider = [outBatch featuresAtIndex:i];
-        whisper_decoder_implOutput * result = [[whisper_decoder_implOutput alloc] initWithVar_1346:(MLMultiArray *)[resultProvider featureValueForName:@"var_1346"].multiArrayValue];
+        whisper_decoder_implOutput * result = [[whisper_decoder_implOutput alloc] initWithCast_76:(MLMultiArray *)[resultProvider featureValueForName:@"cast_76"].multiArrayValue];
         [results addObject:result];
     }
     return results;

data/ext/{src → sources/src}/coreml/whisper-encoder-impl.h

@@ -11,9 +11,8 @@
 
 NS_ASSUME_NONNULL_BEGIN
 
-
 /// Model Prediction Input Type
-API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((visibility("hidden")))
+API_AVAILABLE(macos(10.15), ios(13.0), watchos(6.0), tvos(13.0)) __attribute__((visibility("hidden")))
 @interface whisper_encoder_implInput : NSObject<MLFeatureProvider>
 
 /// logmel_data as 1 × 80 × 3000 3-dimensional array of floats
@@ -23,9 +22,8 @@ API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((v
 
 @end
 
-
 /// Model Prediction Output Type
-API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((visibility("hidden")))
+API_AVAILABLE(macos(10.15), ios(13.0), watchos(6.0), tvos(13.0)) __attribute__((visibility("hidden")))
 @interface whisper_encoder_implOutput : NSObject<MLFeatureProvider>
 
 /// output as multidimensional array of floats
@@ -35,9 +33,8 @@ API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((v
 
 @end
 
-
 /// Class for model loading and prediction
-API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((visibility("hidden")))
+API_AVAILABLE(macos(10.15), ios(13.0), watchos(6.0), tvos(13.0)) __attribute__((visibility("hidden")))
 @interface whisper_encoder_impl : NSObject
 @property (readonly, nonatomic, nullable) MLModel * model;
 
@@ -91,7 +88,7 @@ API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((v
     @param configuration The model configuration
     @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid whisper_encoder_impl instance or NSError object.
 */
-+ (void)loadWithConfiguration:(MLModelConfiguration *)configuration completionHandler:(void (^)(whisper_encoder_impl * _Nullable model, NSError * _Nullable error))handler;
++ (void)loadWithConfiguration:(MLModelConfiguration *)configuration completionHandler:(void (^)(whisper_encoder_impl * _Nullable model, NSError * _Nullable error))handler API_AVAILABLE(macos(11.0), ios(14.0), watchos(7.0), tvos(14.0)) __attribute__((visibility("hidden")));
 
 /**
     Construct whisper_encoder_impl instance asynchronously with URL of .mlmodelc directory and optional configuration.
@@ -102,7 +99,7 @@ API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((v
     @param configuration The model configuration
     @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid whisper_encoder_impl instance or NSError object.
 */
-+ (void)loadContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration completionHandler:(void (^)(whisper_encoder_impl * _Nullable model, NSError * _Nullable error))handler;
++ (void)loadContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration completionHandler:(void (^)(whisper_encoder_impl * _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
@@ -121,9 +118,24 @@ API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((v
 */
 - (nullable whisper_encoder_implOutput *)predictionFromFeatures:(whisper_encoder_implInput *)input options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error;
 
+/**
+    Make an asynchronous prediction using the standard interface
+    @param input an instance of whisper_encoder_implInput to predict from
+    @param completionHandler a block that will be called upon completion of the prediction. error will be nil if no error occurred.
+*/
+- (void)predictionFromFeatures:(whisper_encoder_implInput *)input completionHandler:(void (^)(whisper_encoder_implOutput * _Nullable output, NSError * _Nullable error))completionHandler API_AVAILABLE(macos(14.0), ios(17.0), watchos(10.0), tvos(17.0)) __attribute__((visibility("hidden")));
+
+/**
+    Make an asynchronous prediction using the standard interface
+    @param input an instance of whisper_encoder_implInput to predict from
+    @param options prediction options
+    @param completionHandler a block that will be called upon completion of the prediction. error will be nil if no error occurred.
+*/
+- (void)predictionFromFeatures:(whisper_encoder_implInput *)input options:(MLPredictionOptions *)options completionHandler:(void (^)(whisper_encoder_implOutput * _Nullable output, NSError * _Nullable error))completionHandler API_AVAILABLE(macos(14.0), ios(17.0), watchos(10.0), tvos(17.0)) __attribute__((visibility("hidden")));
+
 /**
     Make a prediction using the convenience interface
-    @param logmel_data as 1 × n_mel × 3000 3-dimensional array of floats:
+    @param logmel_data 1 × 80 × 3000 3-dimensional array of floats
     @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
     @return the prediction as whisper_encoder_implOutput
 */

data/ext/{src → sources/src}/coreml/whisper-encoder-impl.m

@@ -76,10 +76,13 @@
     Such application may want to use `-[MLModel initWithContentsOfURL:configuration:error:]` and `+URLOfModelInThisBundle` to create a MLModel object to pass-in.
 */
 - (instancetype)initWithMLModel:(MLModel *)model {
+    if (model == nil) {
+        return nil;
+    }
     self = [super init];
-    if (!self) { return nil; }
-    _model = model;
-    if (_model == nil) { return nil; }
+    if (self != nil) {
+        _model = model;
+    }
     return self;
 }
 
@@ -176,6 +179,28 @@
     return [[whisper_encoder_implOutput alloc] initWithOutput:(MLMultiArray *)[outFeatures featureValueForName:@"output"].multiArrayValue];
 }
 
+- (void)predictionFromFeatures:(whisper_encoder_implInput *)input completionHandler:(void (^)(whisper_encoder_implOutput * _Nullable output, NSError * _Nullable error))completionHandler {
+    [self.model predictionFromFeatures:input completionHandler:^(id<MLFeatureProvider> prediction, NSError *predictionError) {
+        if (prediction != nil) {
+            whisper_encoder_implOutput *output = [[whisper_encoder_implOutput alloc] initWithOutput:(MLMultiArray *)[prediction featureValueForName:@"output"].multiArrayValue];
+            completionHandler(output, predictionError);
+        } else {
+            completionHandler(nil, predictionError);
+        }
+    }];
+}
+
+- (void)predictionFromFeatures:(whisper_encoder_implInput *)input options:(MLPredictionOptions *)options completionHandler:(void (^)(whisper_encoder_implOutput * _Nullable output, NSError * _Nullable error))completionHandler {
+    [self.model predictionFromFeatures:input options:options completionHandler:^(id<MLFeatureProvider> prediction, NSError *predictionError) {
+        if (prediction != nil) {
+            whisper_encoder_implOutput *output = [[whisper_encoder_implOutput alloc] initWithOutput:(MLMultiArray *)[prediction featureValueForName:@"output"].multiArrayValue];
+            completionHandler(output, predictionError);
+        } else {
+            completionHandler(nil, predictionError);
+        }
+    }];
+}
+
 - (nullable whisper_encoder_implOutput *)predictionFromLogmel_data:(MLMultiArray *)logmel_data error:(NSError * _Nullable __autoreleasing * _Nullable)error {
     whisper_encoder_implInput *input_ = [[whisper_encoder_implInput alloc] initWithLogmel_data:logmel_data];
     return [self predictionFromFeatures:input_ error:error];

data/ext/sources/src/coreml/whisper-encoder.mm

@@ -0,0 +1,73 @@
+#if !__has_feature(objc_arc)
+#error This file must be compiled with automatic reference counting enabled (-fobjc-arc)
+#endif
+
+#import "whisper-encoder.h"
+#import "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];
+
+    // select which device to run the Core ML model on
+    MLModelConfiguration *config = [[MLModelConfiguration alloc] init];
+    // config.computeUnits = MLComputeUnitsCPUAndGPU;
+    //config.computeUnits = MLComputeUnitsCPUAndNeuralEngine;
+    config.computeUnits = MLComputeUnitsAll;
+
+    const void * data = CFBridgingRetain([[whisper_encoder_impl alloc] initWithContentsOfURL:url_model configuration:config 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,
+                             int64_t   n_ctx,
+                             int64_t   n_mel,
+                               float * mel,
+                               float * out) {
+    MLMultiArray * inMultiArray = [
+        [MLMultiArray alloc] initWithDataPointer: mel
+                                           shape: @[@1, @(n_mel), @(n_ctx)]
+                                        dataType: MLMultiArrayDataTypeFloat32
+                                         strides: @[@(n_ctx*n_mel), @(n_ctx), @1]
+                                     deallocator: nil
+                                           error: nil
+    ];
+
+    @autoreleasepool {
+        whisper_encoder_implOutput * outCoreML = [(__bridge id) ctx->data predictionFromLogmel_data:inMultiArray error:nil];
+
+        memcpy(out, outCoreML.output.dataPointer, outCoreML.output.count * sizeof(float));
+    }
+}
+
+#if __cplusplus
+}
+#endif