whisper.rn 0.4.0-rc.4 → 0.4.0-rc.6

package/cpp/whisper.h

@@ -1,6 +1,8 @@
 #ifndef WHISPER_H
 #define WHISPER_H
 
+#include "ggml.h"
+
 #include <stddef.h>
 #include <stdint.h>
 #include <stdbool.h>
@@ -48,7 +50,9 @@ 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");
@@ -76,7 +80,9 @@ 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;
 
     struct whisper_context_params {
         bool  use_gpu;
@@ -108,18 +114,49 @@ 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
-    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);
+    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_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_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),
@@ -279,6 +316,9 @@ 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
@@ -401,6 +441,7 @@ 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
@@ -478,6 +519,11 @@ extern "C" {
         // 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_context_params & whisper_free_params()
@@ -571,8 +617,7 @@ 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);
 
 #ifdef __cplusplus
 }

package/ios/RNWhisper.mm

@@ -142,9 +142,9 @@ RCT_REMAP_METHOD(transcribeFile,
         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:@{
@@ -156,9 +156,9 @@ 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:@{
@@ -279,7 +279,7 @@ 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 +2,6 @@
 
 @interface RNWhisperAudioUtils : NSObject
 
-+ (NSData *)concatShortBuffers:(NSMutableArray<NSValue *> *)buffers sliceNSamples:(NSMutableArray<NSNumber *> *)sliceNSamples;
-+ (void)saveWavFile:(NSData *)rawData audioOutputFile:(NSString *)audioOutputFile;
 + (float *)decodeWaveFile:(NSString*)filePath count:(int *)count;
 
 @end

package/ios/RNWhisperAudioUtils.m

@@ -3,62 +3,6 @@
 
 @implementation RNWhisperAudioUtils
 
-+ (NSData *)concatShortBuffers:(NSMutableArray<NSValue *> *)buffers sliceNSamples:(NSMutableArray<NSNumber *> *)sliceNSamples {
-    NSMutableData *outputData = [NSMutableData data];
-    for (int i = 0; i < buffers.count; i++) {
-        int size = [sliceNSamples objectAtIndex:i].intValue;
-        NSValue *buffer = [buffers objectAtIndex:i];
-        short *bufferPtr = buffer.pointerValue;
-        [outputData appendBytes:bufferPtr length:size * sizeof(short)];
-    }
-    return outputData;
-}
-
-+ (void)saveWavFile:(NSData *)rawData audioOutputFile:(NSString *)audioOutputFile {
-    NSMutableData *outputData = [NSMutableData data];
-    
-    // WAVE header
-    [outputData appendData:[@"RIFF" dataUsingEncoding:NSUTF8StringEncoding]]; // chunk id
-    int chunkSize = CFSwapInt32HostToLittle(36 + rawData.length);
-    [outputData appendBytes:&chunkSize length:sizeof(chunkSize)];
-    [outputData appendData:[@"WAVE" dataUsingEncoding:NSUTF8StringEncoding]]; // format
-    [outputData appendData:[@"fmt " dataUsingEncoding:NSUTF8StringEncoding]]; // subchunk 1 id
-    
-    int subchunk1Size = CFSwapInt32HostToLittle(16);
-    [outputData appendBytes:&subchunk1Size length:sizeof(subchunk1Size)];
-
-    short audioFormat = CFSwapInt16HostToLittle(1); // PCM
-    [outputData appendBytes:&audioFormat length:sizeof(audioFormat)];
-
-    short numChannels = CFSwapInt16HostToLittle(1); // mono
-    [outputData appendBytes:&numChannels length:sizeof(numChannels)];
-
-    int sampleRate = CFSwapInt32HostToLittle(WHISPER_SAMPLE_RATE);
-    [outputData appendBytes:&sampleRate length:sizeof(sampleRate)];
-
-    // (bitDepth * sampleRate * channels) >> 3
-    int byteRate = CFSwapInt32HostToLittle(WHISPER_SAMPLE_RATE * 1 * 16 / 8);
-    [outputData appendBytes:&byteRate length:sizeof(byteRate)];
-
-    // (bitDepth * channels) >> 3
-    short blockAlign = CFSwapInt16HostToLittle(16 / 8);
-    [outputData appendBytes:&blockAlign length:sizeof(blockAlign)];
-
-    // bitDepth
-    short bitsPerSample = CFSwapInt16HostToLittle(16);
-    [outputData appendBytes:&bitsPerSample length:sizeof(bitsPerSample)];
-
-    [outputData appendData:[@"data" dataUsingEncoding:NSUTF8StringEncoding]]; // subchunk 2 id
-    int subchunk2Size = CFSwapInt32HostToLittle((int)rawData.length);
-    [outputData appendBytes:&subchunk2Size length:sizeof(subchunk2Size)];
-
-    // Audio data
-    [outputData appendData:rawData];
-    
-    // Save to file
-    [outputData writeToFile:audioOutputFile atomically:YES];
-}
-
 + (float *)decodeWaveFile:(NSString*)filePath count:(int *)count {
     NSURL *url = [NSURL fileURLWithPath:filePath];
     NSData *fileData = [NSData dataWithContentsOfURL:url];

package/ios/RNWhisperContext.h

@@ -11,29 +11,21 @@
 
 typedef struct {
     __unsafe_unretained id mSelf;
-
-    int jobId;
     NSDictionary* options;
 
+    struct rnwhisper::job * job;
+
     bool isTranscribing;
     bool isRealtime;
     bool isCapturing;
     bool isStoppedByAction;
-    int maxAudioSec;
     int nSamplesTranscribing;
-    NSMutableArray<NSValue *> *shortBufferSlices;
-    NSMutableArray<NSNumber *> *sliceNSamples;
+    std::vector<int> sliceNSamples;
     bool isUseSlices;
     int sliceIndex;
     int transcribeSliceIndex;
-    int audioSliceSec;
     NSString* audioOutputPath;
 
-    bool useVad;
-    int vadMs;
-    float vadThold;
-    float vadFreqThold;
-
     AudioQueueRef queue;
     AudioStreamBasicDescription dataFormat;
     AudioQueueBufferRef buffers[NUM_BUFFERS];

package/ios/RNWhisperContext.mm

@@ -1,5 +1,4 @@
 #import "RNWhisperContext.h"
-#import "RNWhisperAudioUtils.h"
 #import <Metal/Metal.h>
 #include <vector>
 
@@ -95,7 +94,7 @@
     return self->dQueue;
 }
 
-- (void)prepareRealtime:(NSDictionary *)options {
+- (void)prepareRealtime:(int)jobId options:(NSDictionary *)options {
     self->recordState.options = options;
 
     self->recordState.dataFormat.mSampleRate = WHISPER_SAMPLE_RATE; // 16000
@@ -108,74 +107,39 @@
     self->recordState.dataFormat.mReserved = 0;
     self->recordState.dataFormat.mFormatFlags = kLinearPCMFormatFlagIsSignedInteger;
 
-    int maxAudioSecOpt = options[@"realtimeAudioSec"] != nil ? [options[@"realtimeAudioSec"] intValue] : 0;
-    int maxAudioSec = maxAudioSecOpt > 0 ? maxAudioSecOpt : DEFAULT_MAX_AUDIO_SEC;
-    self->recordState.maxAudioSec = maxAudioSec;
-
-    int realtimeAudioSliceSec = options[@"realtimeAudioSliceSec"] != nil ? [options[@"realtimeAudioSliceSec"] intValue] : 0;
-    int audioSliceSec = realtimeAudioSliceSec > 0 && realtimeAudioSliceSec < maxAudioSec ? realtimeAudioSliceSec : maxAudioSec;
-
-    self->recordState.audioOutputPath = options[@"audioOutputPath"];
-
-    self->recordState.useVad = options[@"useVad"] != nil ? [options[@"useVad"] boolValue] : false;
-    self->recordState.vadMs = options[@"vadMs"] != nil ? [options[@"vadMs"] intValue] : 2000;
-    if (self->recordState.vadMs < 2000) self->recordState.vadMs = 2000;
-
-    self->recordState.vadThold = options[@"vadThold"] != nil ? [options[@"vadThold"] floatValue] : 0.6f;
-    self->recordState.vadFreqThold = options[@"vadFreqThold"] != nil ? [options[@"vadFreqThold"] floatValue] : 100.0f;
-
-    self->recordState.audioSliceSec = audioSliceSec;
-    self->recordState.isUseSlices = audioSliceSec < maxAudioSec;
+    self->recordState.isRealtime = true;
+    self->recordState.isTranscribing = false;
+    self->recordState.isCapturing = false;
+    self->recordState.isStoppedByAction = false;
 
     self->recordState.sliceIndex = 0;
     self->recordState.transcribeSliceIndex = 0;
     self->recordState.nSamplesTranscribing = 0;
 
-    [self freeBufferIfNeeded];
-    self->recordState.shortBufferSlices = [NSMutableArray new];
-
-    int16_t *audioBufferI16 = (int16_t *) malloc(audioSliceSec * WHISPER_SAMPLE_RATE * sizeof(int16_t));
-    [self->recordState.shortBufferSlices addObject:[NSValue valueWithPointer:audioBufferI16]];
-
-    self->recordState.sliceNSamples = [NSMutableArray new];
-    [self->recordState.sliceNSamples addObject:[NSNumber numberWithInt:0]];
-
-    self->recordState.isRealtime = true;
-    self->recordState.isTranscribing = false;
-    self->recordState.isCapturing = false;
-    self->recordState.isStoppedByAction = false;
+    self->recordState.sliceNSamples.push_back(0);
+
+    self->recordState.job = rnwhisper::job_new(jobId, [self createParams:options jobId:jobId]);
+    self->recordState.job->set_realtime_params(
+        {
+            .use_vad = options[@"useVad"] != nil ? [options[@"useVad"] boolValue] : false,
+            .vad_ms = options[@"vadMs"] != nil ? [options[@"vadMs"] intValue] : 2000,
+            .vad_thold = options[@"vadThold"] != nil ? [options[@"vadThold"] floatValue] : 0.6f,
+            .freq_thold = options[@"vadFreqThold"] != nil ? [options[@"vadFreqThold"] floatValue] : 100.0f
+        },
+        options[@"realtimeAudioSec"] != nil ? [options[@"realtimeAudioSec"] intValue] : 0,
+        options[@"realtimeAudioSliceSec"] != nil ? [options[@"realtimeAudioSliceSec"] intValue] : 0,
+        options[@"realtimeAudioMinSec"] != nil ? [options[@"realtimeAudioMinSec"] floatValue] : 0,
+        options[@"audioOutputPath"] != nil ? [options[@"audioOutputPath"] UTF8String] : nullptr
+    );
+    self->recordState.isUseSlices = self->recordState.job->audio_slice_sec < self->recordState.job->audio_sec;
 
     self->recordState.mSelf = self;
 }
 
-- (void)freeBufferIfNeeded {
-    if (self->recordState.shortBufferSlices != nil) {
-        for (int i = 0; i < [self->recordState.shortBufferSlices count]; i++) {
-            int16_t *audioBufferI16 = (int16_t *) [self->recordState.shortBufferSlices[i] pointerValue];
-            free(audioBufferI16);
-        }
-        self->recordState.shortBufferSlices = nil;
-    }
-}
-
-bool vad(RNWhisperContextRecordState *state, int16_t* audioBufferI16, int nSamples, int n)
+bool vad(RNWhisperContextRecordState *state, int sliceIndex, int nSamples, int n)
 {
-    bool isSpeech = true;
-    if (!state->isTranscribing && state->useVad) {
-        int sampleSize = (int) (WHISPER_SAMPLE_RATE * state->vadMs / 1000);
-        if (nSamples + n > sampleSize) {
-            int start = nSamples + n - sampleSize;
-            std::vector<float> audioBufferF32Vec(sampleSize);
-            for (int i = 0; i < sampleSize; i++) {
-                audioBufferF32Vec[i] = (float)audioBufferI16[i + start] / 32768.0f;
-            }
-            isSpeech = rn_whisper_vad_simple(audioBufferF32Vec, WHISPER_SAMPLE_RATE, 1000, state->vadThold, state->vadFreqThold, false);
-            NSLog(@"[RNWhisper] VAD result: %d", isSpeech);
-        } else {
-            isSpeech = false;
-        }
-    }
-    return isSpeech;
+    if (state->isTranscribing) return true;
+    return state->job->vad_simple(sliceIndex, nSamples, n);
 }
 
 void AudioInputCallback(void * inUserData,
@@ -196,15 +160,15 @@ void AudioInputCallback(void * inUserData,
     }
 
     int totalNSamples = 0;
-    for (int i = 0; i < [state->sliceNSamples count]; i++) {
-        totalNSamples += [[state->sliceNSamples objectAtIndex:i] intValue];
+    for (int i = 0; i < state->sliceNSamples.size(); i++) {
+        totalNSamples += state->sliceNSamples[i];
     }
 
     const int n = inBuffer->mAudioDataByteSize / 2;
 
-    int nSamples = [state->sliceNSamples[state->sliceIndex] intValue];
+    int nSamples = state->sliceNSamples[state->sliceIndex];
 
-    if (totalNSamples + n > state->maxAudioSec * WHISPER_SAMPLE_RATE) {
+    if (totalNSamples + n > state->job->audio_sec * WHISPER_SAMPLE_RATE) {
         NSLog(@"[RNWhisper] Audio buffer is full, stop capturing");
         state->isCapturing = false;
         [state->mSelf stopAudio];
@@ -218,8 +182,8 @@ void AudioInputCallback(void * inUserData,
             !state->isTranscribing &&
             nSamples != state->nSamplesTranscribing
         ) {
-            int16_t* audioBufferI16 = (int16_t*) [state->shortBufferSlices[state->sliceIndex] pointerValue];
-            if (!vad(state, audioBufferI16, nSamples, 0)) {
+            bool isSamplesEnough = nSamples / WHISPER_SAMPLE_RATE >= state->job->audio_min_sec;
+            if (!isSamplesEnough || !vad(state, state->sliceIndex, nSamples, 0)) {
                 [state->mSelf finishRealtimeTranscribe:state result:@{}];
                 return;
             }
@@ -231,31 +195,25 @@ void AudioInputCallback(void * inUserData,
         return;
     }
 
-    int audioSliceSec = state->audioSliceSec;
-    if (nSamples + n > audioSliceSec * WHISPER_SAMPLE_RATE) {
+    if (nSamples + n > state->job->audio_slice_sec * WHISPER_SAMPLE_RATE) {
         // next slice
         state->sliceIndex++;
         nSamples = 0;
-        int16_t* audioBufferI16 = (int16_t*) malloc(audioSliceSec * WHISPER_SAMPLE_RATE * sizeof(int16_t));
-        [state->shortBufferSlices addObject:[NSValue valueWithPointer:audioBufferI16]];
-        [state->sliceNSamples addObject:[NSNumber numberWithInt:0]];
+        state->sliceNSamples.push_back(0);
     }
 
-    // Append to buffer
     NSLog(@"[RNWhisper] Slice %d has %d samples", state->sliceIndex, nSamples);
 
-    int16_t* audioBufferI16 = (int16_t*) [state->shortBufferSlices[state->sliceIndex] pointerValue];
-    for (int i = 0; i < n; i++) {
-        audioBufferI16[nSamples + i] = ((short*)inBuffer->mAudioData)[i];
-    }
+    state->job->put_pcm_data((short*) inBuffer->mAudioData, state->sliceIndex, nSamples, n);
 
-    bool isSpeech = vad(state, audioBufferI16, nSamples, n);
+    bool isSpeech = vad(state, state->sliceIndex, nSamples, n);
     nSamples += n;
-    state->sliceNSamples[state->sliceIndex] = [NSNumber numberWithInt:nSamples];
+    state->sliceNSamples[state->sliceIndex] = nSamples;
 
     AudioQueueEnqueueBuffer(state->queue, inBuffer, 0, NULL);
 
-    if (!isSpeech) return;
+    bool isSamplesEnough = nSamples / WHISPER_SAMPLE_RATE >= state->job->audio_min_sec;
+    if (!isSamplesEnough || !isSpeech) return;
 
     if (!state->isTranscribing) {
         state->isTranscribing = true;
@@ -267,32 +225,27 @@ void AudioInputCallback(void * inUserData,
 
 - (void)finishRealtimeTranscribe:(RNWhisperContextRecordState*) state result:(NSDictionary*)result {
     // Save wav if needed
-    if (state->audioOutputPath != nil) {
+    if (state->job->audio_output_path != nullptr) {
         // TODO: Append in real time so we don't need to keep all slices & also reduce memory usage
-        [RNWhisperAudioUtils
-            saveWavFile:[RNWhisperAudioUtils concatShortBuffers:state->shortBufferSlices
-                            sliceNSamples:state->sliceNSamples]
-            audioOutputFile:state->audioOutputPath
-        ];
+        rnaudioutils::save_wav_file(
+            rnaudioutils::concat_short_buffers(state->job->pcm_slices, state->sliceNSamples),
+            state->job->audio_output_path
+        );
     }
-    state->transcribeHandler(state->jobId, @"end", result);
+    state->transcribeHandler(state->job->job_id, @"end", result);
+    rnwhisper::job_remove(state->job->job_id);
 }
 
 - (void)fullTranscribeSamples:(RNWhisperContextRecordState*) state {
-    int nSamplesOfIndex = [[state->sliceNSamples objectAtIndex:state->transcribeSliceIndex] intValue];
+    int nSamplesOfIndex = state->sliceNSamples[state->transcribeSliceIndex];
     state->nSamplesTranscribing = nSamplesOfIndex;
     NSLog(@"[RNWhisper] Transcribing %d samples", state->nSamplesTranscribing);
 
-    int16_t* audioBufferI16 = (int16_t*) [state->shortBufferSlices[state->transcribeSliceIndex] pointerValue];
-    float* audioBufferF32 = (float*) malloc(state->nSamplesTranscribing * sizeof(float));
-    // convert I16 to F32
-    for (int i = 0; i < state->nSamplesTranscribing; i++) {
-        audioBufferF32[i] = (float)audioBufferI16[i] / 32768.0f;
-    }
+    float* pcmf32 = state->job->pcm_slice_to_f32(state->transcribeSliceIndex, state->nSamplesTranscribing);
+
     CFTimeInterval timeStart = CACurrentMediaTime();
-    struct whisper_full_params params = [state->mSelf getParams:state->options jobId:state->jobId];
-    int code = [state->mSelf fullTranscribe:state->jobId params:params audioData:audioBufferF32 audioDataCount:state->nSamplesTranscribing];
-    free(audioBufferF32);
+    int code = [state->mSelf fullTranscribe:state->job audioData:pcmf32 audioDataCount:state->nSamplesTranscribing];
+    free(pcmf32);
     CFTimeInterval timeEnd = CACurrentMediaTime();
     const float timeRecording = (float) state->nSamplesTranscribing / (float) state->dataFormat.mSampleRate;
 
@@ -312,7 +265,7 @@ void AudioInputCallback(void * inUserData,
         result[@"error"] = [NSString stringWithFormat:@"Transcribe failed with code %d", code];
     }
 
-    nSamplesOfIndex = [[state->sliceNSamples objectAtIndex:state->transcribeSliceIndex] intValue];
+    nSamplesOfIndex = state->sliceNSamples[state->transcribeSliceIndex];
 
     bool isStopped = state->isStoppedByAction || (
         !state->isCapturing &&
@@ -340,10 +293,10 @@ void AudioInputCallback(void * inUserData,
         [state->mSelf finishRealtimeTranscribe:state result:result];
     } else if (code == 0) {
         result[@"isCapturing"] = @(true);
-        state->transcribeHandler(state->jobId, @"transcribe", result);
+        state->transcribeHandler(state->job->job_id, @"transcribe", result);
     } else {
         result[@"isCapturing"] = @(true);
-        state->transcribeHandler(state->jobId, @"transcribe", result);
+        state->transcribeHandler(state->job->job_id, @"transcribe", result);
     }
 
     if (continueNeeded) {
@@ -371,8 +324,7 @@ void AudioInputCallback(void * inUserData,
     onTranscribe:(void (^)(int, NSString *, NSDictionary *))onTranscribe
 {
     self->recordState.transcribeHandler = onTranscribe;
-    self->recordState.jobId = jobId;
-    [self prepareRealtime:options];
+    [self prepareRealtime:jobId options:options];
 
     OSStatus status = AudioQueueNewInput(
         &self->recordState.dataFormat,
@@ -413,9 +365,9 @@ struct rnwhisper_segments_callback_data {
     dispatch_async(dQueue, ^{
         self->recordState.isStoppedByAction = false;
         self->recordState.isTranscribing = true;
-        self->recordState.jobId = jobId;
 
-        whisper_full_params params = [self getParams:options jobId:jobId];
+        whisper_full_params params = [self createParams:options jobId:jobId];
+
         if (options[@"onProgress"] && [options[@"onProgress"] boolValue]) {
             params.progress_callback = [](struct whisper_context * /*ctx*/, struct whisper_state * /*state*/, int progress, void * user_data) {
                 void (^onProgress)(int) = (__bridge void (^)(int))user_data;
@@ -460,8 +412,10 @@ struct rnwhisper_segments_callback_data {
             };
             params.new_segment_callback_user_data = &user_data;
         }
-        int code = [self fullTranscribe:jobId params:params audioData:audioData audioDataCount:audioDataCount];
-        self->recordState.jobId = -1;
+
+        rnwhisper::job* job = rnwhisper::job_new(jobId, params);;
+        int code = [self fullTranscribe:job audioData:audioData audioDataCount:audioDataCount];
+        rnwhisper::job_remove(jobId);
         self->recordState.isTranscribing = false;
         onEnd(code);
     });
@@ -476,7 +430,7 @@ struct rnwhisper_segments_callback_data {
 }
 
 - (void)stopTranscribe:(int)jobId {
-    rn_whisper_abort_transcribe(jobId);
+    if (self->recordState.job) self->recordState.job->abort();
     if (self->recordState.isRealtime && self->recordState.isCapturing) {
         [self stopAudio];
         if (!self->recordState.isTranscribing) {
@@ -490,13 +444,11 @@ struct rnwhisper_segments_callback_data {
 }
 
 - (void)stopCurrentTranscribe {
-    if (!self->recordState.jobId) {
-        return;
-    }
-    [self stopTranscribe:self->recordState.jobId];
+    if (self->recordState.job == nullptr) return;
+    [self stopTranscribe:self->recordState.job->job_id];
 }
 
-- (struct whisper_full_params)getParams:(NSDictionary *)options jobId:(int)jobId {
+- (struct whisper_full_params)createParams:(NSDictionary *)options jobId:(int)jobId {
     struct whisper_full_params params = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
 
     const int n_threads = options[@"maxThreads"] != nil ?
@@ -517,7 +469,7 @@ struct rnwhisper_segments_callback_data {
     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 ? [options[@"language"] UTF8String] : "auto";
+    params.language         = options[@"language"] != nil ? strdup([options[@"language"] UTF8String]) : "auto";
     params.n_threads        = n_threads > 0 ? n_threads : default_n_threads;
     params.offset_ms        = 0;
     params.no_context       = true;
@@ -534,7 +486,6 @@ struct rnwhisper_segments_callback_data {
     if (options[@"maxContext"] != nil) {
         params.n_max_text_ctx = [options[@"maxContext"] intValue];
     }
-
     if (options[@"offset"] != nil) {
         params.offset_ms = [options[@"offset"] intValue];
     }
@@ -550,39 +501,20 @@ struct rnwhisper_segments_callback_data {
     if (options[@"temperatureInc"] != nil) {
         params.temperature_inc = [options[@"temperature_inc"] floatValue];
     }
-
     if (options[@"prompt"] != nil) {
-        params.initial_prompt = [options[@"prompt"] UTF8String];
+        params.initial_prompt = strdup([options[@"prompt"] UTF8String]);
     }
 
-    // abort handler
-    bool *abort_ptr = rn_whisper_assign_abort_map(jobId);
-    params.encoder_begin_callback = [](struct whisper_context * /*ctx*/, struct whisper_state * /*state*/, void * user_data) {
-        bool is_aborted = *(bool*)user_data;
-        return !is_aborted;
-    };
-    params.encoder_begin_callback_user_data = abort_ptr;
-    params.abort_callback = [](void * user_data) {
-        bool is_aborted = *(bool*)user_data;
-        return is_aborted;
-    };
-    params.abort_callback_user_data = abort_ptr;
-
     return params;
 }
 
-- (int)fullTranscribe:(int)jobId
-  params:(struct whisper_full_params)params
+- (int)fullTranscribe:(rnwhisper::job *)job
   audioData:(float *)audioData
   audioDataCount:(int)audioDataCount
 {
     whisper_reset_timings(self->ctx);
-
-    int code = whisper_full(self->ctx, params, audioData, audioDataCount);
-    if (rn_whisper_transcribe_is_aborted(jobId)) {
-        code = -999;
-    }
-    rn_whisper_remove_abort_map(jobId);
+    int code = whisper_full(self->ctx, job->params, audioData, audioDataCount);
+    if (job && job->is_aborted()) code = -999;
     // if (code == 0) {
     //     whisper_print_timings(self->ctx);
     // }
@@ -616,7 +548,6 @@ struct rnwhisper_segments_callback_data {
 - (void)invalidate {
     [self stopCurrentTranscribe];
     whisper_free(self->ctx);
-    [self freeBufferIfNeeded];
 }
 
 @end