whisper.rn 0.4.0-rc.9 → 0.4.0

package/README.md

@@ -25,6 +25,8 @@ npm install whisper.rn
 
 Please re-run `npx pod-install` again.
 
+By default, `whisper.rn` will use pre-built `rnwhisper.xcframework` for iOS. If you want to build from source, please set `RNWHISPER_BUILD_FROM_SOURCE` to `1` in your Podfile.
+
 If you want to use `medium` or `large` model, the [Extended Virtual Addressing](https://developer.apple.com/documentation/bundleresources/entitlements/com_apple_developer_kernel_extended-virtual-addressing) capability is recommended to enable on iOS project.
 
 #### Android
@@ -36,7 +38,9 @@ Add proguard rule if it's enabled in project (android/app/proguard-rules.pro):
 -keep class com.rnwhisper.** { *; }
 ```
 
-For build, it's recommended to use `ndkVersion = "24.0.8215888"` (or above) in your root project build configuration for Apple Silicon Macs. Otherwise please follow this trobleshooting [issue](./TROUBLESHOOTING.md#android-got-build-error-unknown-host-cpu-architecture-arm64-on-apple-silicon-macs).
+By default, `whisper.rn` will use pre-built libraries for Android. If you want to build from source, please set `rnwhisperBuildFromSource` to `true` in `android/gradle.properties`.
+
+For build from source, it's recommended to use `ndkVersion = "24.0.8215888"` (or above) in your root project build configuration for Apple Silicon Macs. Otherwise please follow this trobleshooting [issue](./TROUBLESHOOTING.md#android-got-build-error-unknown-host-cpu-architecture-arm64-on-apple-silicon-macs).
 
 #### Expo
 

package/android/build.gradle

@@ -53,9 +53,18 @@ android {
       }
     }
   }
-  externalNativeBuild {
-    cmake {
-      path = file('src/main/CMakeLists.txt')
+  def rnwhisperBuildFromSource = project.properties["rnwhisperBuildFromSource"]
+  if (rnwhisperBuildFromSource == "true") {
+    externalNativeBuild {
+      cmake {
+        path = file('src/main/CMakeLists.txt')
+      }
+    }
+    // Exclude jniLibs
+    sourceSets {
+      main {
+        jniLibs.srcDirs = []
+      }
     }
   }
   buildTypes {

package/android/src/main/CMakeLists.txt

@@ -2,16 +2,35 @@ cmake_minimum_required(VERSION 3.10)
 
 project(whisper.rn)
 
-set(CMAKE_CXX_STANDARD 11)
+set(CMAKE_CXX_STANDARD 17)
 set(RNWHISPER_LIB_DIR ${CMAKE_SOURCE_DIR}/../../../cpp)
 
+include_directories(
+    ${RNWHISPER_LIB_DIR}
+    ${RNWHISPER_LIB_DIR}/ggml-cpu
+)
+
 set(
     SOURCE_FILES
     ${RNWHISPER_LIB_DIR}/ggml.c
     ${RNWHISPER_LIB_DIR}/ggml-alloc.c
     ${RNWHISPER_LIB_DIR}/ggml-backend.cpp
+    ${RNWHISPER_LIB_DIR}/ggml-backend-reg.cpp
+    ${RNWHISPER_LIB_DIR}/ggml-cpu/amx/amx.cpp
+    ${RNWHISPER_LIB_DIR}/ggml-cpu/amx/mmq.cpp
+    ${RNWHISPER_LIB_DIR}/ggml-cpu/ggml-cpu.c
+    ${RNWHISPER_LIB_DIR}/ggml-cpu/ggml-cpu.cpp
+    ${RNWHISPER_LIB_DIR}/ggml-cpu/quants.c
+    ${RNWHISPER_LIB_DIR}/ggml-cpu/traits.cpp
+    ${RNWHISPER_LIB_DIR}/ggml-cpu/repack.cpp
+    ${RNWHISPER_LIB_DIR}/ggml-cpu/unary-ops.cpp
+    ${RNWHISPER_LIB_DIR}/ggml-cpu/binary-ops.cpp
+    ${RNWHISPER_LIB_DIR}/ggml-cpu/vec.cpp
+    ${RNWHISPER_LIB_DIR}/ggml-cpu/ops.cpp
+    ${RNWHISPER_LIB_DIR}/ggml-opt.cpp
+    ${RNWHISPER_LIB_DIR}/ggml-threading.cpp
     ${RNWHISPER_LIB_DIR}/ggml-quants.c
-    ${RNWHISPER_LIB_DIR}/ggml-aarch64.c
+    ${RNWHISPER_LIB_DIR}/gguf.cpp
     ${RNWHISPER_LIB_DIR}/whisper.cpp
     ${RNWHISPER_LIB_DIR}/rn-audioutils.cpp
     ${RNWHISPER_LIB_DIR}/rn-whisper.cpp
@@ -20,45 +39,56 @@ set(
 
 find_library(LOG_LIB log)
 
-function(build_library target_name)
+function(build_library target_name arch cpu_flags)
+    if (NOT ${arch} STREQUAL "generic")
+        set(SOURCE_FILES_ARCH
+            ${RNWHISPER_LIB_DIR}/ggml-cpu/arch/${arch}/quants.c
+            ${RNWHISPER_LIB_DIR}/ggml-cpu/arch/${arch}/repack.cpp
+        )
+    endif ()
+
     add_library(
         ${target_name}
         SHARED
         ${SOURCE_FILES}
+        ${SOURCE_FILES_ARCH}
     )
 
     target_link_libraries(${target_name} ${LOG_LIB} android)
 
-    if (${target_name} STREQUAL "whisper_v8fp16_va")
-        target_compile_options(${target_name} PRIVATE -march=armv8.2-a+fp16)
-    elseif (${target_name} STREQUAL "whisper_vfpv4")
-        target_compile_options(${target_name} PRIVATE -mfpu=neon-vfpv4)
+    if (${arch} STREQUAL "generic")
+        target_compile_options(${target_name} PRIVATE -DWSP_GGML_CPU_GENERIC)
     endif ()
 
+    target_compile_options(${target_name} PRIVATE -DWSP_GGML_USE_CPU -DWSP_GGML_USE_CPU_REPACK -pthread ${cpu_flags})
+
     if (${CMAKE_BUILD_TYPE} STREQUAL "Debug")
         target_compile_options(${target_name} PRIVATE -DRNWHISPER_ANDROID_ENABLE_LOGGING)
     endif ()
 
     # NOTE: If you want to debug the native code, you can uncomment if and endif
+    # Note that it will be extremely slow
     # if (NOT ${CMAKE_BUILD_TYPE} STREQUAL "Debug")
-
-        target_compile_options(${target_name} PRIVATE -O3 -DNDEBUG -pthread)
+        target_compile_options(${target_name} PRIVATE -O3 -DNDEBUG)
         target_compile_options(${target_name} PRIVATE -fvisibility=hidden -fvisibility-inlines-hidden)
         target_compile_options(${target_name} PRIVATE -ffunction-sections -fdata-sections)
 
         target_link_options(${target_name} PRIVATE -Wl,--gc-sections)
         target_link_options(${target_name} PRIVATE -Wl,--exclude-libs,ALL)
         target_link_options(${target_name} PRIVATE -flto)
-
     # endif ()
 endfunction()
 
-build_library("whisper") # Default target
+build_library("rnwhisper" "generic" "")
 
 if (${ANDROID_ABI} STREQUAL "arm64-v8a")
-    build_library("whisper_v8fp16_va")
+    build_library("rnwhisper_v8fp16_va_2" "arm" "-march=armv8.2-a+fp16")
 elseif (${ANDROID_ABI} STREQUAL "armeabi-v7a")
-    build_library("whisper_vfpv4")
+    build_library("rnwhisper_vfpv4" "arm" "-mfpu=neon-vfpv4")
+elseif (${ANDROID_ABI} STREQUAL "x86_64")
+    # x86_64 target
+    build_library("rnwhisper_x86_64" "x86" "-march=x86-64" "-mtune=intel" "-msse4.2" "-mpopcnt")
 endif ()
 
+
 include_directories(${RNWHISPER_LIB_DIR})

package/android/src/main/java/com/rnwhisper/WhisperContext.java

@@ -27,6 +27,8 @@ import java.io.PushbackInputStream;
 public class WhisperContext {
   public static final String NAME = "RNWhisperContext";
 
+  private static String loadedLibrary = "";
+
   private static final int SAMPLE_RATE = 16000;
   private static final int CHANNEL_CONFIG = AudioFormat.CHANNEL_IN_MONO;
   private static final int AUDIO_FORMAT = AudioFormat.ENCODING_PCM_16BIT;
@@ -433,64 +435,60 @@ public class WhisperContext {
 
   static {
     Log.d(NAME, "Primary ABI: " + Build.SUPPORTED_ABIS[0]);
-    boolean loadVfpv4 = false;
-    boolean loadV8fp16 = false;
-    if (isArmeabiV7a()) {
-      // armeabi-v7a needs runtime detection support
-      String cpuInfo = cpuInfo();
-      if (cpuInfo != null) {
-        Log.d(NAME, "CPU info: " + cpuInfo);
-        if (cpuInfo.contains("vfpv4")) {
-          Log.d(NAME, "CPU supports vfpv4");
-          loadVfpv4 = true;
-        }
-      }
-    } else if (isArmeabiV8a()) {
-      // ARMv8.2a needs runtime detection support
-      String cpuInfo = cpuInfo();
-      if (cpuInfo != null) {
-        Log.d(NAME, "CPU info: " + cpuInfo);
-        if (cpuInfo.contains("fphp")) {
-          Log.d(NAME, "CPU supports fp16 arithmetic");
-          loadV8fp16 = true;
-        }
-      }
-    }
 
-    if (loadVfpv4) {
-      Log.d(NAME, "Loading libwhisper_vfpv4.so");
-      System.loadLibrary("whisper_vfpv4");
-    } else if (loadV8fp16) {
-      Log.d(NAME, "Loading libwhisper_v8fp16_va.so");
-      System.loadLibrary("whisper_v8fp16_va");
+    String cpuFeatures = WhisperContext.getCpuFeatures();
+    Log.d(NAME, "CPU features: " + cpuFeatures);
+    boolean hasFp16 = cpuFeatures.contains("fp16") || cpuFeatures.contains("fphp");
+    Log.d(NAME, "- hasFp16: " + hasFp16);
+
+    if (WhisperContext.isArm64V8a()) {
+      if (hasFp16) {
+        Log.d(NAME, "Loading librnwhisper_v8fp16_va_2.so");
+        System.loadLibrary("rnwhisper_v8fp16_va_2");
+        loadedLibrary = "rnwhisper_v8fp16_va_2";
+      }
+    } else if (WhisperContext.isArmeabiV7a()) {
+      Log.d(NAME, "Loading librnwhisper_vfpv4.so");
+      System.loadLibrary("rnwhisper_vfpv4");
+      loadedLibrary = "rnwhisper_vfpv4";
+    } else if (WhisperContext.isX86_64()) {
+      Log.d(NAME, "Loading librnwhisper_x86_64.so");
+      System.loadLibrary("rnwhisper_x86_64");
+      loadedLibrary = "rnwhisper_x86_64";
     } else {
-      Log.d(NAME, "Loading libwhisper.so");
-      System.loadLibrary("whisper");
+      Log.d(NAME, "ARM32 is not supported, skipping loading library");
     }
   }
 
+  private static boolean isArm64V8a() {
+    return Build.SUPPORTED_ABIS[0].equals("arm64-v8a");
+  }
+
   private static boolean isArmeabiV7a() {
     return Build.SUPPORTED_ABIS[0].equals("armeabi-v7a");
   }
 
-  private static boolean isArmeabiV8a() {
-    return Build.SUPPORTED_ABIS[0].equals("arm64-v8a");
+  private static boolean isX86_64() {
+    return Build.SUPPORTED_ABIS[0].equals("x86_64");
   }
 
-  private static String cpuInfo() {
+  private static String getCpuFeatures() {
     File file = new File("/proc/cpuinfo");
     StringBuilder stringBuilder = new StringBuilder();
     try {
       BufferedReader bufferedReader = new BufferedReader(new FileReader(file));
       String line;
       while ((line = bufferedReader.readLine()) != null) {
+        if (line.startsWith("Features")) {
           stringBuilder.append(line);
+          break;
+        }
       }
       bufferedReader.close();
       return stringBuilder.toString();
     } catch (IOException e) {
       Log.w(NAME, "Couldn't read /proc/cpuinfo", e);
-      return null;
+      return "";
     }
   }
 

package/android/src/main/jni.cpp

@@ -155,7 +155,10 @@ Java_com_rnwhisper_WhisperContext_initContext(
         JNIEnv *env, jobject thiz, jstring model_path_str) {
     UNUSED(thiz);
     struct whisper_context_params cparams;
+
+    // TODO: Expose dtw_token_timestamps and dtw_aheads_preset
     cparams.dtw_token_timestamps = false;
+    // cparams.dtw_aheads_preset = WHISPER_AHEADS_BASE;
 
     struct whisper_context *context = nullptr;
     const char *model_path_chars = env->GetStringUTFChars(model_path_str, nullptr);
@@ -173,7 +176,10 @@ Java_com_rnwhisper_WhisperContext_initContextWithAsset(
 ) {
     UNUSED(thiz);
     struct whisper_context_params cparams;
+
+    // TODO: Expose dtw_token_timestamps and dtw_aheads_preset
     cparams.dtw_token_timestamps = false;
+    // cparams.dtw_aheads_preset = WHISPER_AHEADS_BASE;
 
     struct whisper_context *context = nullptr;
     const char *model_path_chars = env->GetStringUTFChars(model_path_str, nullptr);
@@ -190,7 +196,10 @@ Java_com_rnwhisper_WhisperContext_initContextWithInputStream(
 ) {
     UNUSED(thiz);
     struct whisper_context_params cparams;
+
+    // TODO: Expose dtw_token_timestamps and dtw_aheads_preset
     cparams.dtw_token_timestamps = false;
+    // cparams.dtw_aheads_preset = WHISPER_AHEADS_BASE;
 
     struct whisper_context *context = nullptr;
     context = whisper_init_from_input_stream(env, input_stream, cparams);

package/cpp/coreml/whisper-compat.h

@@ -0,0 +1,10 @@
+#import <CoreML/CoreML.h>
+
+@interface MLModel (Compat)
+- (void) predictionFromFeatures:(id<MLFeatureProvider>) input
+              completionHandler:(void (^)(id<MLFeatureProvider> output, NSError * error)) completionHandler;
+
+- (void) predictionFromFeatures:(id<MLFeatureProvider>) input
+                        options:(MLPredictionOptions *) options
+              completionHandler:(void (^)(id<MLFeatureProvider> output, NSError * error)) completionHandler;
+@end

package/cpp/coreml/whisper-compat.m

@@ -0,0 +1,35 @@
+#import "whisper-compat.h"
+#import <Foundation/Foundation.h>
+
+@implementation MLModel (Compat)
+
+#if !defined(MAC_OS_X_VERSION_14_00) || MAC_OS_X_VERSION_MAX_ALLOWED < MAC_OS_X_VERSION_14_00
+
+- (void) predictionFromFeatures:(id<MLFeatureProvider>) input
+              completionHandler:(void (^)(id<MLFeatureProvider> output, NSError * error)) completionHandler {
+    [NSOperationQueue.new addOperationWithBlock:^{
+        NSError *error = nil;
+        id<MLFeatureProvider> prediction = [self predictionFromFeatures:input error:&error];
+
+        [NSOperationQueue.mainQueue addOperationWithBlock:^{
+            completionHandler(prediction, error);
+        }];
+    }];
+}
+
+- (void) predictionFromFeatures:(id<MLFeatureProvider>) input
+                        options:(MLPredictionOptions *) options
+              completionHandler:(void (^)(id<MLFeatureProvider> output, NSError * error)) completionHandler {
+    [NSOperationQueue.new addOperationWithBlock:^{
+        NSError *error = nil;
+        id<MLFeatureProvider> prediction = [self predictionFromFeatures:input options:options error:&error];
+
+        [NSOperationQueue.mainQueue addOperationWithBlock:^{
+            completionHandler(prediction, error);
+        }];
+    }];
+}
+
+#endif
+
+@end

package/cpp/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
 */

package/cpp/coreml/whisper-decoder-impl.m

@@ -8,6 +8,7 @@
 #error This file must be compiled with automatic reference counting enabled (-fobjc-arc)
 #endif
 
+#import "whisper-compat.h"
 #import "whisper-decoder-impl.h"
 
 @implementation whisper_decoder_implInput
@@ -39,21 +40,21 @@
 
 @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 +81,13 @@
     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 +181,29 @@
 - (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 +218,7 @@
     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;

package/cpp/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
 */