whisper.rn 0.4.0-rc.8 → 0.4.0-rc.9

package/cpp/whisper.h

@@ -84,9 +84,48 @@ extern "C" {
     typedef int32_t whisper_token;
     typedef int32_t whisper_seq_id;
 
+    enum whisper_alignment_heads_preset {
+        WHISPER_AHEADS_NONE,
+        WHISPER_AHEADS_N_TOP_MOST,  // All heads from the N-top-most text-layers
+        WHISPER_AHEADS_CUSTOM,
+        WHISPER_AHEADS_TINY_EN,
+        WHISPER_AHEADS_TINY,
+        WHISPER_AHEADS_BASE_EN,
+        WHISPER_AHEADS_BASE,
+        WHISPER_AHEADS_SMALL_EN,
+        WHISPER_AHEADS_SMALL,
+        WHISPER_AHEADS_MEDIUM_EN,
+        WHISPER_AHEADS_MEDIUM,
+        WHISPER_AHEADS_LARGE_V1,
+        WHISPER_AHEADS_LARGE_V2,
+        WHISPER_AHEADS_LARGE_V3,
+        WHISPER_AHEADS_LARGE_V3_TURBO,
+    };
+
+    typedef struct whisper_ahead {
+        int n_text_layer;
+        int n_head;
+    } whisper_ahead;
+
+    typedef struct whisper_aheads {
+        size_t n_heads;
+        const whisper_ahead * heads;
+    } whisper_aheads;
+
     struct whisper_context_params {
         bool  use_gpu;
         bool  use_coreml;
+        bool  flash_attn;
+        int   gpu_device;  // CUDA device
+
+        // [EXPERIMENTAL] Token-level timestamps with DTW
+        bool dtw_token_timestamps;
+        enum whisper_alignment_heads_preset dtw_aheads_preset;
+
+        int dtw_n_top;
+        struct whisper_aheads dtw_aheads;
+
+        size_t dtw_mem_size; // TODO: remove
     };
 
     typedef struct whisper_token_data {
@@ -103,6 +142,11 @@ extern "C" {
         int64_t t0;        // start time of the token
         int64_t t1;        //   end time of the token
 
+        // [EXPERIMENTAL] Token-level timestamps with DTW
+        // do not use if you haven't computed token-level timestamps with dtw
+        // Roughly corresponds to the moment in audio in which the token was output
+        int64_t t_dtw;
+
         float vlen;        // voice length of the token
     } whisper_token_data;
 
@@ -196,6 +240,13 @@ extern "C" {
     //                     GPU, by caching compiled 'blobs' there.
     //                     Set to nullptr if not used.
     // Returns 0 on success. If OpenVINO is not enabled in build, this simply returns 1.
+    WHISPER_API int whisper_ctx_init_openvino_encoder_with_state(
+        struct whisper_context * ctx,
+          struct whisper_state * state,
+                    const char * model_path,
+                    const char * device,
+                    const char * cache_dir);
+
     WHISPER_API int whisper_ctx_init_openvino_encoder(
         struct whisper_context * ctx,
                     const char * model_path,
@@ -224,22 +275,6 @@ extern "C" {
                                int   n_samples,
                                int   n_threads);
 
-    // Convert RAW PCM audio to log mel spectrogram but applies a Phase Vocoder to speed up the audio x2.
-    // The resulting spectrogram is stored inside the default state of the provided whisper context.
-    // Returns 0 on success
-    WHISPER_API int whisper_pcm_to_mel_phase_vocoder(
-        struct whisper_context * ctx,
-                   const float * samples,
-                           int   n_samples,
-                           int   n_threads);
-
-    WHISPER_API int whisper_pcm_to_mel_phase_vocoder_with_state(
-        struct whisper_context * ctx,
-          struct whisper_state * state,
-                   const float * samples,
-                           int   n_samples,
-                           int   n_threads);
-
     // This can be used to set a custom log mel spectrogram inside the default state of the provided whisper context.
     // Use this instead of whisper_pcm_to_mel() if you want to provide your own log mel spectrogram.
     // n_mel must be 80
@@ -296,7 +331,7 @@ extern "C" {
     // 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 -1 on failure
+    // Returns a negative number on failure - the number of tokens that would have been returned
     // TODO: not sure if correct
     WHISPER_API int whisper_tokenize(
             struct whisper_context * ctx,
@@ -304,8 +339,12 @@ extern "C" {
                      whisper_token * tokens,
                                int   n_max_tokens);
 
+    // Return the number of tokens in the provided text
+    // Equivalent to: -whisper_tokenize(ctx, text, NULL, 0)
+    int whisper_token_count(struct whisper_context * ctx, const char * text);
+
     // Largest language id (i.e. number of available languages - 1)
-    WHISPER_API int whisper_lang_max_id();
+    WHISPER_API int whisper_lang_max_id(void);
 
     // Return the id of the specified language, returns -1 if not found
     // Examples:
@@ -385,6 +424,24 @@ extern "C" {
     WHISPER_API whisper_token whisper_token_transcribe(struct whisper_context * ctx);
 
     // Performance information from the default state.
+    struct whisper_timings {
+        int64_t load_us;
+        int64_t t_start_us;
+        int32_t fail_p;
+        int32_t fail_h;
+        int64_t t_mel_us;
+        int32_t n_sample;
+        int32_t n_encode;
+        int32_t n_decode;
+        int32_t n_batchd;
+        int32_t n_prompt;
+        int64_t t_sample_us;
+        int64_t t_encode_us;
+        int64_t t_decode_us;
+        int64_t t_batchd_us;
+        int64_t t_prompt_us;
+    };
+    WHISPER_API struct whisper_timings * whisper_get_timings(struct whisper_context * ctx);
     WHISPER_API void whisper_print_timings(struct whisper_context * ctx);
     WHISPER_API void whisper_reset_timings(struct whisper_context * ctx);
 
@@ -412,11 +469,6 @@ extern "C" {
     // If it returns false, the computation is aborted
     typedef bool (*whisper_encoder_begin_callback)(struct whisper_context * ctx, struct whisper_state * state, void * user_data);
 
-    // Abort callback
-    // If not NULL, called before ggml computation
-    // If it returns true, the computation is aborted
-    typedef bool (*whisper_abort_callback)(void * user_data);
-
     // Logits filter callback
     // Can be used to modify the logits before sampling
     // If not NULL, called after applying temperature to logits
@@ -458,15 +510,19 @@ extern "C" {
 
         // [EXPERIMENTAL] speed-up techniques
         // note: these can significantly reduce the quality of the output
-        bool speed_up;          // speed-up the audio by 2x using Phase Vocoder
         bool debug_mode;        // enable debug_mode provides extra info (eg. Dump log_mel)
         int  audio_ctx;         // overwrite the audio context size (0 = use default)
 
         // [EXPERIMENTAL] [TDRZ] tinydiarize
         bool tdrz_enable;       // enable tinydiarize speaker turn detection
 
+        // A regular expression that matches tokens to suppress
+        const char * suppress_regex;
+
         // tokens to provide to the whisper decoder as initial prompt
         // these are prepended to any existing text context from a previous call
+        // use whisper_tokenize() to convert text to tokens
+        // maximum of whisper_n_text_ctx()/2 tokens are used (typically 224)
         const char * initial_prompt;
         const whisper_token * prompt_tokens;
         int prompt_n_tokens;
@@ -513,7 +569,7 @@ extern "C" {
         void * encoder_begin_callback_user_data;
 
         // called each time before ggml computation starts
-        whisper_abort_callback abort_callback;
+        wsp_ggml_abort_callback abort_callback;
         void * abort_callback_user_data;
 
         // called by each decoder to filter obtained logits
@@ -527,10 +583,10 @@ extern "C" {
     };
 
     // 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();
-    WHISPER_API struct whisper_context_params whisper_context_default_params(void);
+    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);
+    WHISPER_API struct whisper_full_params   whisper_full_default_params       (enum whisper_sampling_strategy strategy);
 
     // Run the entire model: PCM -> log mel spectrogram -> encoder -> decoder -> text
     // Not thread safe for same context

package/ios/RNWhisper.mm

@@ -50,6 +50,7 @@ RCT_REMAP_METHOD(initContext,
     BOOL isBundleAsset = [[modelOptions objectForKey:@"isBundleAsset"] boolValue];
     BOOL useGpu = [[modelOptions objectForKey:@"useGpu"] boolValue];
     BOOL useCoreMLIos = [[modelOptions objectForKey:@"useCoreMLIos"] boolValue];
+    BOOL useFlashAttn = [[modelOptions objectForKey:@"useFlashAttn"] boolValue];
 
     // For support debug assets in development mode
     BOOL downloadCoreMLAssets = [[modelOptions objectForKey:@"downloadCoreMLAssets"] boolValue];
@@ -79,6 +80,7 @@ RCT_REMAP_METHOD(initContext,
         contextId:contextId
         noCoreML:!useCoreMLIos
         noMetal:!useGpu
+        useFlashAttn:useFlashAttn
     ];
     if ([context getContext] == NULL) {
         reject(@"whisper_cpp_error", @"Failed to load the model", nil);
@@ -103,42 +105,17 @@ RCT_REMAP_METHOD(initContext,
   ];
 }
 
-RCT_REMAP_METHOD(transcribeFile,
-                 withContextId:(int)contextId
-                 withJobId:(int)jobId
-                 withWaveFile:(NSString *)waveFilePath
-                 withOptions:(NSDictionary *)options
-                 withResolver:(RCTPromiseResolveBlock)resolve
-                 withRejecter:(RCTPromiseRejectBlock)reject)
+- (void)transcribeData:(RNWhisperContext *)context
+    withContextId:(int)contextId
+    withJobId:(int)jobId
+    withData:(float *)data
+    withDataCount:(int)count
+    withOptions:(NSDictionary *)options
+    withResolver:(RCTPromiseResolveBlock)resolve
+    withRejecter:(RCTPromiseRejectBlock)reject
 {
-    RNWhisperContext *context = contexts[[NSNumber numberWithInt:contextId]];
-
-    if (context == nil) {
-        reject(@"whisper_error", @"Context not found", nil);
-        return;
-    }
-    if ([context isCapturing]) {
-        reject(@"whisper_error", @"The context is in realtime transcribe mode", nil);
-        return;
-    }
-    if ([context isTranscribing]) {
-        reject(@"whisper_error", @"Context is already transcribing", nil);
-        return;
-    }
-
-    NSString *path = waveFilePath;
-    if ([path hasPrefix:@"http://"] || [path hasPrefix:@"https://"]) {
-        path = [RNWhisperDownloader downloadFile:path toFile:nil];
-    }
-
-    int count = 0;
-    float *waveFile = [RNWhisperAudioUtils decodeWaveFile:path count:&count];
-    if (waveFile == nil) {
-        reject(@"whisper_error", @"Invalid file", nil);
-        return;
-    }
-    [context transcribeFile:jobId
-        audioData:waveFile
+    [context transcribeData:jobId
+        audioData:data
         audioDataCount:count
         options:options
         onProgress: ^(int progress) {
@@ -171,11 +148,9 @@ RCT_REMAP_METHOD(transcribeFile,
         }
         onEnd: ^(int code) {
             if (code != 0 && code != 999) {
-                free(waveFile);
                 reject(@"whisper_cpp_error", [NSString stringWithFormat:@"Failed to transcribe the file. Code: %d", code], nil);
                 return;
             }
-            free(waveFile);
             NSMutableDictionary *result = [context getTextSegments];
             result[@"isAborted"] = @([context isStoppedByAction]);
             resolve(result);
@@ -183,6 +158,99 @@ RCT_REMAP_METHOD(transcribeFile,
     ];
 }
 
+RCT_REMAP_METHOD(transcribeFile,
+                 withContextId:(int)contextId
+                 withJobId:(int)jobId
+                 withWaveFile:(NSString *)waveFilePathOrDataBase64
+                 withOptions:(NSDictionary *)options
+                 withResolver:(RCTPromiseResolveBlock)resolve
+                 withRejecter:(RCTPromiseRejectBlock)reject)
+{
+    RNWhisperContext *context = contexts[[NSNumber numberWithInt:contextId]];
+
+    if (context == nil) {
+        reject(@"whisper_error", @"Context not found", nil);
+        return;
+    }
+    if ([context isCapturing]) {
+        reject(@"whisper_error", @"The context is in realtime transcribe mode", nil);
+        return;
+    }
+    if ([context isTranscribing]) {
+        reject(@"whisper_error", @"Context is already transcribing", nil);
+        return;
+    }
+
+    float *data = nil;
+    int count = 0;
+    if ([waveFilePathOrDataBase64 hasPrefix:@"http://"] || [waveFilePathOrDataBase64 hasPrefix:@"https://"]) {
+        NSString *path = [RNWhisperDownloader downloadFile:waveFilePathOrDataBase64 toFile:nil];
+        data = [RNWhisperAudioUtils decodeWaveFile:path count:&count];
+    } else if ([waveFilePathOrDataBase64 hasPrefix:@"data:audio/wav;base64,"]) {
+        NSData *waveData = [[NSData alloc] initWithBase64EncodedString:[waveFilePathOrDataBase64 substringFromIndex:22] options:0];
+        data = [RNWhisperAudioUtils decodeWaveData:waveData count:&count cutHeader:YES];
+    } else {
+        data = [RNWhisperAudioUtils decodeWaveFile:waveFilePathOrDataBase64 count:&count];
+    }
+    if (data == nil) {
+        reject(@"whisper_error", @"Invalid file", nil);
+        return;
+    }
+
+    [self transcribeData:context
+        withContextId:contextId
+        withJobId:jobId
+        withData:data
+        withDataCount:count
+        withOptions:options
+        withResolver:resolve
+        withRejecter:reject
+    ];
+}
+
+RCT_REMAP_METHOD(transcribeData,
+                 withContextId:(int)contextId
+                 withJobId:(int)jobId
+                 withData:(NSString *)dataBase64 // pcm data base64 encoded
+                 withOptions:(NSDictionary *)options
+                 withResolver:(RCTPromiseResolveBlock)resolve
+                 withRejecter:(RCTPromiseRejectBlock)reject)
+{
+  RNWhisperContext *context = contexts[[NSNumber numberWithInt:contextId]];
+
+  if (context == nil) {
+      reject(@"whisper_error", @"Context not found", nil);
+      return;
+  }
+  if ([context isCapturing]) {
+      reject(@"whisper_error", @"The context is in realtime transcribe mode", nil);
+      return;
+  }
+  if ([context isTranscribing]) {
+      reject(@"whisper_error", @"Context is already transcribing", nil);
+      return;
+  }
+
+  NSData *pcmData = [[NSData alloc] initWithBase64EncodedString:dataBase64 options:0];
+  int count = 0;
+  float *data = [RNWhisperAudioUtils decodeWaveData:pcmData count:&count cutHeader:NO];
+
+  if (data == nil) {
+      reject(@"whisper_error", @"Invalid data", nil);
+      return;
+  }
+
+  [self transcribeData:context
+      withContextId:contextId
+      withJobId:jobId
+      withData:data
+      withDataCount:count
+      withOptions:options
+      withResolver:resolve
+      withRejecter:reject
+  ];
+}
+
 RCT_REMAP_METHOD(startRealtimeTranscribe,
                  withContextId:(int)contextId
                  withJobId:(int)jobId
@@ -244,6 +312,25 @@ RCT_REMAP_METHOD(abortTranscribe,
     resolve(nil);
 }
 
+RCT_REMAP_METHOD(bench,
+                 withContextId:(int)contextId
+                 withMaxThreads:(int)maxThreads
+                 withResolver:(RCTPromiseResolveBlock)resolve
+                 withRejecter:(RCTPromiseRejectBlock)reject)
+{
+    RNWhisperContext *context = contexts[[NSNumber numberWithInt:contextId]];
+    if (context == nil) {
+        reject(@"whisper_error", @"Context not found", nil);
+        return;
+    }
+    if ([context isTranscribing]) {
+        reject(@"whisper_error", @"The context is transcribing", nil);
+        return;
+    }
+    NSString *result = [context bench:maxThreads];
+    resolve(result);
+}
+
 RCT_REMAP_METHOD(releaseContext,
                  withContextId:(int)contextId
                  withResolver:(RCTPromiseResolveBlock)resolve

package/ios/RNWhisperAudioUtils.h

@@ -2,6 +2,7 @@
 
 @interface RNWhisperAudioUtils : NSObject
 
++ (float *)decodeWaveData:(NSData*)data count:(int *)count cutHeader:(BOOL)cutHeader;
 + (float *)decodeWaveFile:(NSString*)filePath count:(int *)count;
 
 @end

package/ios/RNWhisperAudioUtils.m

@@ -3,25 +3,32 @@
 
 @implementation RNWhisperAudioUtils
 
++ (float *)decodeWaveData:(NSData*)data count:(int *)count cutHeader:(BOOL)cutHeader {
+  NSData *waveData = data;
+  if (cutHeader) {
+    // just cut 44 bytes from the beginning
+    waveData = [data subdataWithRange:NSMakeRange(44, [data length]-44)];
+  }
+  const short *shortArray = (const short *)[waveData bytes];
+  int shortCount = (int) ([waveData length] / sizeof(short));
+  float *floatArray = (float *) malloc(shortCount * sizeof(float));
+  for (NSInteger i = 0; i < shortCount; i++) {
+      float floatValue = ((float)shortArray[i]) / 32767.0;
+      floatValue = MAX(floatValue, -1.0);
+      floatValue = MIN(floatValue, 1.0);
+      floatArray[i] = floatValue;
+  }
+  *count = shortCount;
+  return floatArray;
+}
+
 + (float *)decodeWaveFile:(NSString*)filePath count:(int *)count {
     NSURL *url = [NSURL fileURLWithPath:filePath];
     NSData *fileData = [NSData dataWithContentsOfURL:url];
     if (fileData == nil) {
         return nil;
     }
-    NSMutableData *waveData = [[NSMutableData alloc] init];
-    [waveData appendData:[fileData subdataWithRange:NSMakeRange(44, [fileData length]-44)]];
-    const short *shortArray = (const short *)[waveData bytes];
-    int shortCount = (int) ([waveData length] / sizeof(short));
-    float *floatArray = (float *) malloc(shortCount * sizeof(float));
-    for (NSInteger i = 0; i < shortCount; i++) {
-        float floatValue = ((float)shortArray[i]) / 32767.0;
-        floatValue = MAX(floatValue, -1.0);
-        floatValue = MIN(floatValue, 1.0);
-        floatArray[i] = floatValue;
-    }
-    *count = shortCount;
-    return floatArray;
+    return [RNWhisperAudioUtils decodeWaveData:fileData count:count cutHeader:YES];
 }
 
 @end

package/ios/RNWhisperContext.h

@@ -42,7 +42,7 @@ typedef struct {
     bool isMetalEnabled;
 }
 
-+ (instancetype)initWithModelPath:(NSString *)modelPath contextId:(int)contextId noCoreML:(BOOL)noCoreML noMetal:(BOOL)noMetal;
++ (instancetype)initWithModelPath:(NSString *)modelPath contextId:(int)contextId noCoreML:(BOOL)noCoreML noMetal:(BOOL)noMetal useFlashAttn:(BOOL)useFlashAttn;
 - (bool)isMetalEnabled;
 - (NSString *)reasonNoMetal;
 - (struct whisper_context *)getContext;
@@ -50,7 +50,7 @@ typedef struct {
 - (OSStatus)transcribeRealtime:(int)jobId
     options:(NSDictionary *)options
     onTranscribe:(void (^)(int, NSString *, NSDictionary *))onTranscribe;
-- (void)transcribeFile:(int)jobId
+- (void)transcribeData:(int)jobId
     audioData:(float *)audioData
     audioDataCount:(int)audioDataCount
     options:(NSDictionary *)options
@@ -63,6 +63,7 @@ typedef struct {
 - (bool)isTranscribing;
 - (bool)isStoppedByAction;
 - (NSMutableDictionary *)getTextSegments;
+- (NSString *)bench:(int)maxThreads;
 - (void)invalidate;
 
 @end

package/ios/RNWhisperContext.mm

@@ -10,12 +10,17 @@
     contextId:(int)contextId
     noCoreML:(BOOL)noCoreML
     noMetal:(BOOL)noMetal
+    useFlashAttn:(BOOL)useFlashAttn
 {
     RNWhisperContext *context = [[RNWhisperContext alloc] init];
     context->contextId = contextId;
     struct whisper_context_params cparams;
     NSString *reasonNoMetal = @"";
     cparams.use_gpu = !noMetal;
+    cparams.flash_attn = useFlashAttn;
+
+    // TODO: Figure out why it leads to re-init crash
+    cparams.dtw_token_timestamps = false;
 
     cparams.use_coreml = !noCoreML;
 #ifndef WHISPER_USE_COREML
@@ -353,9 +358,10 @@ void AudioInputCallback(void * inUserData,
 struct rnwhisper_segments_callback_data {
     void (^onNewSegments)(NSDictionary *);
     int total_n_new;
+    bool tdrzEnable;
 };
 
-- (void)transcribeFile:(int)jobId
+- (void)transcribeData:(int)jobId
     audioData:(float *)audioData
     audioDataCount:(int)audioDataCount
     options:(NSDictionary *)options
@@ -386,12 +392,18 @@ struct rnwhisper_segments_callback_data {
                 NSMutableArray *segments = [[NSMutableArray alloc] init];
                 for (int i = data->total_n_new - n_new; i < data->total_n_new; i++) {
                     const char * text_cur = whisper_full_get_segment_text(ctx, i);
-                    text = [text stringByAppendingString:[NSString stringWithUTF8String:text_cur]];
+                    NSMutableString *mutable_ns_text = [NSMutableString stringWithUTF8String:text_cur];
+
+                    if (data->tdrzEnable && whisper_full_get_segment_speaker_turn_next(ctx, i)) {
+                        [mutable_ns_text appendString:@" [SPEAKER_TURN]"];
+                    }
+
+                    text = [text stringByAppendingString:mutable_ns_text];
 
                     const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
                     const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
                     NSDictionary *segment = @{
-                        @"text": [NSString stringWithUTF8String:text_cur],
+                        @"text": [NSString stringWithString:mutable_ns_text],
                         @"t0": [NSNumber numberWithLongLong:t0],
                         @"t1": [NSNumber numberWithLongLong:t1]
                     };
@@ -409,7 +421,8 @@ struct rnwhisper_segments_callback_data {
             };
             struct rnwhisper_segments_callback_data user_data = {
                 .onNewSegments = onNewSegments,
-                .total_n_new = 0
+                .tdrzEnable = options[@"tdrzEnable"] && [options[@"tdrzEnable"] boolValue],
+                .total_n_new = 0,
             };
             params.new_segment_callback_user_data = &user_data;
         }
@@ -469,7 +482,6 @@ struct rnwhisper_segments_callback_data {
     params.print_progress   = false;
     params.print_timestamps = false;
     params.print_special    = false;
-    params.speed_up         = options[@"speedUp"] != nil ? [options[@"speedUp"] boolValue] : false;
     params.translate        = options[@"translate"] != nil ? [options[@"translate"] boolValue] : false;
     params.language         = options[@"language"] != nil ? strdup([options[@"language"] UTF8String]) : "auto";
     params.n_threads        = n_threads > 0 ? n_threads : default_n_threads;
@@ -481,6 +493,7 @@ struct rnwhisper_segments_callback_data {
         params.max_len = [options[@"maxLen"] intValue];
     }
     params.token_timestamps = options[@"tokenTimestamps"] != nil ? [options[@"tokenTimestamps"] boolValue] : false;
+    params.tdrz_enable = options[@"tdrzEnable"] != nil ? [options[@"tdrzEnable"] boolValue] : false;
 
     if (options[@"bestOf"] != nil) {
         params.greedy.best_of = [options[@"bestOf"] intValue];
@@ -530,12 +543,21 @@ struct rnwhisper_segments_callback_data {
     NSMutableArray *segments = [[NSMutableArray alloc] init];
     for (int i = 0; i < n_segments; i++) {
         const char * text_cur = whisper_full_get_segment_text(self->ctx, i);
-        text = [text stringByAppendingString:[NSString stringWithUTF8String:text_cur]];
+        NSMutableString *mutable_ns_text = [NSMutableString stringWithUTF8String:text_cur];
+
+        // Simplified condition
+        if (self->recordState.options[@"tdrzEnable"] &&
+            [self->recordState.options[@"tdrzEnable"] boolValue] &&
+            whisper_full_get_segment_speaker_turn_next(self->ctx, i)) {
+            [mutable_ns_text appendString:@" [SPEAKER_TURN]"];
+        }
+
+        text = [text stringByAppendingString:mutable_ns_text];
 
         const int64_t t0 = whisper_full_get_segment_t0(self->ctx, i);
         const int64_t t1 = whisper_full_get_segment_t1(self->ctx, i);
         NSDictionary *segment = @{
-            @"text": [NSString stringWithUTF8String:text_cur],
+            @"text": [NSString stringWithString:mutable_ns_text],
             @"t0": [NSNumber numberWithLongLong:t0],
             @"t1": [NSNumber numberWithLongLong:t1]
         };
@@ -547,6 +569,16 @@ struct rnwhisper_segments_callback_data {
     return result;
 }
 
+- (NSString *)bench:(int)maxThreads {
+    const int n_threads = maxThreads > 0 ? maxThreads : 0;
+
+    const int max_threads = (int) [[NSProcessInfo processInfo] processorCount];
+    // Use 2 threads by default on 4-core devices, 4 threads on more cores
+    const int default_n_threads = max_threads == 4 ? 2 : MIN(4, max_threads);
+    NSString *result = [NSString stringWithUTF8String:rnwhisper::bench(self->ctx, n_threads).c_str()];
+    return result;
+}
+
 - (void)invalidate {
     [self stopCurrentTranscribe];
     whisper_free(self->ctx);

package/jest/mock.js

@@ -45,11 +45,19 @@ if (!NativeModules.RNWhisper) {
         })
       })
     }),
+    bench: jest.fn(() => Promise.resolve({
+      config: 'NEON',
+      nThreads: 1,
+      encodeMs: 1,
+      decodeMs: 1,
+      batchMs: 1,
+      promptMs: 1,
+    })),
     releaseContext: jest.fn(() => Promise.resolve()),
     releaseAllContexts: jest.fn(() => Promise.resolve()),
 
     // iOS AudioSession utils
-    getAudioSessionCurrentCategory: jest.fn(() => Promise.resolve({ 
+    getAudioSessionCurrentCategory: jest.fn(() => Promise.resolve({
       category: 'AVAudioSessionCategoryPlayAndRecord',
       options: [],
     })),

package/lib/commonjs/NativeRNWhisper.js.map

@@ -1 +1 @@
-{"version":3,"names":["_reactNative","require","_default","TurboModuleRegistry","get","exports","default"],"sourceRoot":"../../src","sources":["NativeRNWhisper.ts"],"mappings":";;;;;;AACA,IAAAA,YAAA,GAAAC,OAAA;AAAkD,IAAAC,QAAA,GAiGnCC,gCAAmB,CAACC,GAAG,CAAO,WAAW,CAAC;AAAAC,OAAA,CAAAC,OAAA,GAAAJ,QAAA"}
+{"version":3,"names":["_reactNative","require","_default","TurboModuleRegistry","get","exports","default"],"sourceRoot":"../../src","sources":["NativeRNWhisper.ts"],"mappings":";;;;;;AACA,IAAAA,YAAA,GAAAC,OAAA;AAAkD,IAAAC,QAAA,GA0GnCC,gCAAmB,CAACC,GAAG,CAAO,WAAW,CAAC;AAAAC,OAAA,CAAAC,OAAA,GAAAJ,QAAA"}