whisper.rn 0.4.0-rc.1 → 0.4.0-rc.11

package/cpp/whisper.h

@@ -1,10 +1,21 @@
 #ifndef WHISPER_H
 #define WHISPER_H
 
+#include "ggml.h"
+#include "ggml-cpu.h"
+
 #include <stddef.h>
 #include <stdint.h>
 #include <stdbool.h>
 
+#ifdef __GNUC__
+#    define WHISPER_DEPRECATED(func, hint) func __attribute__((deprecated(hint)))
+#elif defined(_MSC_VER)
+#    define WHISPER_DEPRECATED(func, hint) __declspec(deprecated(hint)) func
+#else
+#    define WHISPER_DEPRECATED(func, hint) func
+#endif
+
 #ifdef WHISPER_SHARED
 #    ifdef _WIN32
 #        ifdef WHISPER_BUILD
@@ -21,7 +32,6 @@
 
 #define WHISPER_SAMPLE_RATE 16000
 #define WHISPER_N_FFT       400
-#define WHISPER_N_MEL       80
 #define WHISPER_HOP_LENGTH  160
 #define WHISPER_CHUNK_SIZE  30
 
@@ -41,7 +51,9 @@ extern "C" {
     //
     //     ...
     //
-    //     struct whisper_context * ctx = whisper_init_from_file("/path/to/ggml-base.en.bin");
+    //     whisper_context_params cparams = whisper_context_default_params();
+    //
+    //     struct whisper_context * ctx = whisper_init_from_file_with_params("/path/to/ggml-base.en.bin", cparams);
     //
     //     if (whisper_full(ctx, wparams, pcmf32.data(), pcmf32.size()) != 0) {
     //         fprintf(stderr, "failed to process audio\n");
@@ -69,7 +81,53 @@ extern "C" {
     struct whisper_state;
     struct whisper_full_params;
 
-    typedef int whisper_token;
+    typedef int32_t whisper_pos;
+    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 {
         whisper_token id;  // token id
@@ -85,6 +143,11 @@ 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;
 
@@ -96,21 +159,74 @@ extern "C" {
         void  (*close)(void * ctx);
     } whisper_model_loader;
 
+    // grammar element type
+    enum whisper_gretype {
+        // end of rule definition
+        WHISPER_GRETYPE_END            = 0,
+
+        // start of alternate definition for rule
+        WHISPER_GRETYPE_ALT            = 1,
+
+        // non-terminal element: reference to rule
+        WHISPER_GRETYPE_RULE_REF       = 2,
+
+        // terminal element: character (code point)
+        WHISPER_GRETYPE_CHAR           = 3,
+
+        // inverse char(s) ([^a], [^a-b] [^abc])
+        WHISPER_GRETYPE_CHAR_NOT       = 4,
+
+        // modifies a preceding WHISPER_GRETYPE_CHAR or LLAMA_GRETYPE_CHAR_ALT to
+        // be an inclusive range ([a-z])
+        WHISPER_GRETYPE_CHAR_RNG_UPPER = 5,
+
+        // modifies a preceding WHISPER_GRETYPE_CHAR or
+        // WHISPER_GRETYPE_CHAR_RNG_UPPER to add an alternate char to match ([ab], [a-zA])
+        WHISPER_GRETYPE_CHAR_ALT       = 6,
+    };
+
+    typedef struct whisper_grammar_element {
+        enum whisper_gretype type;
+        uint32_t             value; // Unicode code point or rule ID
+    } whisper_grammar_element;
+
     // Various functions for loading a ggml whisper model.
     // Allocate (almost) all memory needed for the model.
     // Return NULL on failure
-#ifdef WHISPER_USE_COREML
-    WHISPER_API struct whisper_context * whisper_init_from_file_no_coreml(const char * path_model);
-#endif
-    WHISPER_API struct whisper_context * whisper_init_from_file(const char * path_model);
-    WHISPER_API struct whisper_context * whisper_init_from_buffer(void * buffer, size_t buffer_size);
-    WHISPER_API struct whisper_context * whisper_init(struct whisper_model_loader * loader);
+    WHISPER_API struct whisper_context * whisper_init_from_file_with_params  (const char * path_model,              struct whisper_context_params params);
+    WHISPER_API struct whisper_context * whisper_init_from_buffer_with_params(void * buffer, size_t buffer_size,    struct whisper_context_params params);
+    WHISPER_API struct whisper_context * whisper_init_with_params            (struct whisper_model_loader * loader, struct whisper_context_params params);
 
     // These are the same as the above, but the internal state of the context is not allocated automatically
     // It is the responsibility of the caller to allocate the state using whisper_init_state() (#523)
-    WHISPER_API struct whisper_context * whisper_init_from_file_no_state(const char * path_model);
-    WHISPER_API struct whisper_context * whisper_init_from_buffer_no_state(void * buffer, size_t buffer_size);
-    WHISPER_API struct whisper_context * whisper_init_no_state(struct whisper_model_loader * loader);
+    WHISPER_API struct whisper_context * whisper_init_from_file_with_params_no_state  (const char * path_model,              struct whisper_context_params params);
+    WHISPER_API struct whisper_context * whisper_init_from_buffer_with_params_no_state(void * buffer, size_t buffer_size,    struct whisper_context_params params);
+    WHISPER_API struct whisper_context * whisper_init_with_params_no_state            (struct whisper_model_loader * loader, struct whisper_context_params params);
+
+    WHISPER_DEPRECATED(
+        WHISPER_API struct whisper_context * whisper_init_from_file(const char * path_model),
+        "use whisper_init_from_file_with_params instead"
+    );
+    WHISPER_DEPRECATED(
+        WHISPER_API struct whisper_context * whisper_init_from_buffer(void * buffer, size_t buffer_size),
+        "use whisper_init_from_buffer_with_params instead"
+    );
+    WHISPER_DEPRECATED(
+        WHISPER_API struct whisper_context * whisper_init(struct whisper_model_loader * loader),
+        "use whisper_init_with_params instead"
+    );
+    WHISPER_DEPRECATED(
+        WHISPER_API struct whisper_context * whisper_init_from_file_no_state(const char * path_model),
+        "use whisper_init_from_file_with_params_no_state instead"
+    );
+    WHISPER_DEPRECATED(
+        WHISPER_API struct whisper_context * whisper_init_from_buffer_no_state(void * buffer, size_t buffer_size),
+        "use whisper_init_from_buffer_with_params_no_state instead"
+    );
+    WHISPER_DEPRECATED(
+        WHISPER_API struct whisper_context * whisper_init_no_state(struct whisper_model_loader * loader),
+        "use whisper_init_with_params_no_state instead"
+    );
 
     WHISPER_API struct whisper_state * whisper_init_state(struct whisper_context * ctx);
 
@@ -125,6 +241,13 @@ 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,
@@ -135,6 +258,7 @@ extern "C" {
     WHISPER_API void whisper_free      (struct whisper_context * ctx);
     WHISPER_API void whisper_free_state(struct whisper_state * state);
     WHISPER_API void whisper_free_params(struct whisper_full_params * params);
+    WHISPER_API void whisper_free_context_params(struct whisper_context_params * params);
 
     // Convert RAW PCM audio to log mel spectrogram.
     // The resulting spectrogram is stored inside the default state of the provided whisper context.
@@ -152,22 +276,6 @@ 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
@@ -224,7 +332,7 @@ 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,
@@ -232,8 +340,12 @@ 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:
@@ -244,6 +356,9 @@ extern "C" {
     // Return the short string of the specified language id (e.g. 2 -> "de"), returns nullptr if not found
     WHISPER_API const char * whisper_lang_str(int id);
 
+    // Return the short string of the specified language name (e.g. 2 -> "german"), returns nullptr if not found
+    WHISPER_API const char * whisper_lang_str_full(int id);
+
     // Use mel data at offset_ms to try and auto-detect the spoken language
     // Make sure to call whisper_pcm_to_mel() or whisper_set_mel() first
     // Returns the top language id or negative on failure
@@ -310,6 +425,14 @@ extern "C" {
     WHISPER_API whisper_token whisper_token_transcribe(struct whisper_context * ctx);
 
     // Performance information from the default state.
+    struct whisper_timings {
+        float sample_ms;
+        float encode_ms;
+        float decode_ms;
+        float batchd_ms;
+        float prompt_ms;
+    };
+    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);
 
@@ -337,11 +460,6 @@ 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
@@ -366,6 +484,7 @@ extern "C" {
 
         bool translate;
         bool no_context;        // do not use past transcription (if any) as initial prompt for the decoder
+        bool no_timestamps;     // do not generate timestamps
         bool single_segment;    // force single segment output (useful for streaming)
         bool print_special;     // print special tokens (e.g. <SOT>, <EOT>, <BEG>, etc.)
         bool print_progress;    // print progress information
@@ -382,15 +501,19 @@ 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;
@@ -400,8 +523,8 @@ 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
@@ -412,7 +535,7 @@ extern "C" {
         float temperature_inc;
         float entropy_thold;    // similar to OpenAI's "compression_ratio_threshold"
         float logprob_thold;
-        float no_speech_thold;  // TODO: not implemented
+        float no_speech_thold;
 
         struct {
             int best_of;    // ref: https://github.com/openai/whisper/blob/f82bc59f5ea234d4b97fb2860842ed38519f7e65/whisper/transcribe.py#L264
@@ -437,17 +560,24 @@ 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
         whisper_logits_filter_callback logits_filter_callback;
         void * logits_filter_callback_user_data;
+
+        const whisper_grammar_element ** grammar_rules;
+        size_t                           n_grammar_rules;
+        size_t                           i_start_rule;
+        float                            grammar_penalty;
     };
 
-    // NOTE: this function allocates memory, and it is the responsibility of the caller to free the pointer - see whisper_free_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);
+    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
@@ -534,9 +664,11 @@ extern "C" {
 
     // Control logging output; default behavior is to print to stderr
 
-    typedef void (*whisper_log_callback)(const char * line);
-    WHISPER_API void whisper_set_log_callback(whisper_log_callback callback);
+    WHISPER_API void whisper_log_set(wsp_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

package/ios/RNWhisper.mm

@@ -48,7 +48,9 @@ RCT_REMAP_METHOD(initContext,
 
     NSString *modelPath = [modelOptions objectForKey:@"filePath"];
     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];
@@ -77,6 +79,8 @@ RCT_REMAP_METHOD(initContext,
         initWithModelPath:path
         contextId:contextId
         noCoreML:!useCoreMLIos
+        noMetal:!useGpu
+        useFlashAttn:useFlashAttn
     ];
     if ([context getContext] == NULL) {
         reject(@"whisper_cpp_error", @"Failed to load the model", nil);
@@ -85,7 +89,11 @@ RCT_REMAP_METHOD(initContext,
 
     [contexts setObject:context forKey:[NSNumber numberWithInt:contextId]];
 
-    resolve([NSNumber numberWithInt:contextId]);
+    resolve(@{
+        @"contextId": @(contextId),
+        @"gpu": @([context isMetalEnabled]),
+        @"reasonNoGPU": [context reasonNoMetal],
+    });
 }
 
 - (NSArray *)supportedEvents {
@@ -97,48 +105,23 @@ 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) {
-            if (rn_whisper_transcribe_is_aborted(jobId)) {
-                return;
-            }
+            rnwhisper::job* job = rnwhisper::job_get(jobId);
+            if (job && job->is_aborted()) return;
+
             dispatch_async(dispatch_get_main_queue(), ^{
                 [self sendEventWithName:@"@RNWhisper_onTranscribeProgress"
                     body:@{
@@ -150,9 +133,9 @@ RCT_REMAP_METHOD(transcribeFile,
             });
         }
         onNewSegments: ^(NSDictionary *result) {
-            if (rn_whisper_transcribe_is_aborted(jobId)) {
-                return;
-            }
+            rnwhisper::job* job = rnwhisper::job_get(jobId);
+            if (job && job->is_aborted()) return;
+
             dispatch_async(dispatch_get_main_queue(), ^{
                 [self sendEventWithName:@"@RNWhisper_onTranscribeNewSegments"
                     body:@{
@@ -164,12 +147,10 @@ RCT_REMAP_METHOD(transcribeFile,
             });
         }
         onEnd: ^(int code) {
-            if (code != 0) {
-                free(waveFile);
+            if (code != 0 && code != 999) {
                 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);
@@ -177,6 +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
@@ -238,6 +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
@@ -273,7 +366,7 @@ RCT_REMAP_METHOD(releaseAllContexts,
         [context invalidate];
     }
 
-    rn_whisper_abort_all_transcribe(); // graceful abort
+    rnwhisper::job_abort_all(); // graceful abort
 
     [contexts removeAllObjects];
     contexts = nil;

package/ios/RNWhisperAudioUtils.h

@@ -2,8 +2,7 @@
 
 @interface RNWhisperAudioUtils : NSObject
 
-+ (NSData *)concatShortBuffers:(NSMutableArray<NSValue *> *)buffers sliceNSamples:(NSMutableArray<NSNumber *> *)sliceNSamples;
-+ (void)saveWavFile:(NSData *)rawData audioOutputFile:(NSString *)audioOutputFile;
++ (float *)decodeWaveData:(NSData*)data count:(int *)count cutHeader:(BOOL)cutHeader;
 + (float *)decodeWaveFile:(NSString*)filePath count:(int *)count;
 
 @end