feat: add realtimeAudioMinSec option

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

@@ -110,6 +110,9 @@ public int startRealtimeTranscribe(int jobId, ReadableMap options) {
     final int audioSliceSec = realtimeAudioSliceSec > 0 && realtimeAudioSliceSec < audioSec ? realtimeAudioSliceSec : audioSec;
     isUseSlices = audioSliceSec < audioSec;
 
+    double realtimeAudioMinSec = options.hasKey("realtimeAudioMinSec") ? options.getDouble("realtimeAudioMinSec") : 0;
+    final double audioMinSec = realtimeAudioMinSec > 0.5 && realtimeAudioMinSec <= audioSliceSec ? realtimeAudioMinSec : 1;
+
     createRealtimeTranscribeJob(jobId, context, options);
 
     sliceNSamples = new ArrayList<Integer>();
@@ -144,7 +147,8 @@ public void run() {
                 ) {
                   finishRealtimeTranscribe(Arguments.createMap());
                 } else if (!isTranscribing) {
-                  if (!vad(sliceIndex, nSamples, 0)) {
+                  boolean isSamplesEnough = nSamples / SAMPLE_RATE >= audioMinSec;
+                  if (!isSamplesEnough || !vad(sliceIndex, nSamples, 0)) {
                     finishRealtimeTranscribe(Arguments.createMap());
                     break;
                   }
@@ -169,7 +173,8 @@ public void run() {
               nSamples += n;
               sliceNSamples.set(sliceIndex, nSamples);
 
-              if (!isSpeech) continue;
+              boolean isSamplesEnough = nSamples / SAMPLE_RATE >= audioMinSec;
+              if (!isSamplesEnough || !isSpeech) continue;
 
               if (!isTranscribing && nSamples > SAMPLE_RATE / 2) {
                 isTranscribing = true;

android/src/main/jni.cpp

@@ -199,7 +199,7 @@ struct whisper_full_params createFullParams(JNIEnv *env, jobject options) {
     params.print_progress = false;
     params.print_timestamps = false;
     params.print_special = false;
-    
+
     int max_threads = std::thread::hardware_concurrency();
     // Use 2 threads by default on 4-core devices, 4 threads on more cores
     int default_n_threads = max_threads == 4 ? 2 : min(4, max_threads);
@@ -307,7 +307,7 @@ Java_com_rnwhisper_WhisperContext_fullWithNewJob(
         // whisper_print_timings(context);
     }
     env->ReleaseFloatArrayElements(audio_data, audio_data_arr, JNI_ABORT);
-    
+
     if (job->is_aborted()) code = -999;
     rnwhisper::job_remove(job_id);
     return code;
@@ -339,6 +339,7 @@ Java_com_rnwhisper_WhisperContext_createRealtimeTranscribeJob(
         vad,
         readablemap::getInt(env, options, "realtimeAudioSec", 0),
         readablemap::getInt(env, options, "realtimeAudioSliceSec", 0),
+        readablemap::getFloat(env, options, "realtimeAudioMinSec", 0),
         audio_output_path_str
     );
 }

cpp/rn-whisper.cpp

@@ -63,12 +63,14 @@ void job::set_realtime_params(
     vad_params params,
     int sec,
     int slice_sec,
+    float min_sec,
     const char* output_path
 ) {
     vad = params;
     if (vad.vad_ms < 2000) vad.vad_ms = 2000;
     audio_sec = sec > 0 ? sec : DEFAULT_MAX_AUDIO_SEC;
     audio_slice_sec = slice_sec > 0 && slice_sec < audio_sec ? slice_sec : audio_sec;
+    audio_min_sec = min_sec >= 0.5 && min_sec <= audio_slice_sec ? min_sec : 1.0f;
     audio_output_path = output_path;
 }
 

cpp/rn-whisper.h

@@ -31,9 +31,10 @@ struct job {
     vad_params vad;
     int audio_sec = 0;
     int audio_slice_sec = 0;
+    float audio_min_sec = 0;
     const char* audio_output_path = nullptr;
     std::vector<short *> pcm_slices;
-    void set_realtime_params(vad_params vad, int sec, int slice_sec, const char* output_path);
+    void set_realtime_params(vad_params vad, int sec, int slice_sec, float min_sec, const char* output_path);
     bool vad_simple(int slice_index, int n_samples, int n);
     void put_pcm_data(short* pcm, int slice_index, int n_samples, int n);
     float* pcm_slice_to_f32(int slice_index, int size);
@@ -46,4 +47,4 @@ job* job_get(int job_id);
 
 } // namespace rnwhisper
 
-#endif // RNWHISPER_H
+#endif // RNWHISPER_H

ios/RNWhisperContext.mm

@@ -128,6 +128,7 @@ - (void)prepareRealtime:(int)jobId options:(NSDictionary *)options {
         },
         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;
@@ -181,7 +182,8 @@ void AudioInputCallback(void * inUserData,
             !state->isTranscribing &&
             nSamples != state->nSamplesTranscribing
         ) {
-            if (!vad(state, state->sliceIndex, 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;
             }
@@ -210,7 +212,8 @@ void AudioInputCallback(void * inUserData,
 
     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;

src/index.ts

@@ -94,6 +94,11 @@ export type TranscribeRealtimeOptions = TranscribeOptions & {
    * (Default: Equal to `realtimeMaxAudioSec`)
    */
   realtimeAudioSliceSec?: number
+  /**
+   * Min duration of audio to start transcribe in seconds for each slice.
+   * The minimum value is 0.5 ms and maximum value is realtimeAudioSliceSec (Default: 1)
+   */
+  realtimeAudioMinSec?: number
   /**
    * Output path for audio file. If not set, the audio file will not be saved
    * (Default: Undefined)