update

transcribe/pipelines/pipe_vad.py

@@ -28,12 +28,12 @@ class VadPipe(BasePipe):
     def init(cls):
         if cls.vac is None:
             cls.vac = FixedVADIterator(
-                threshold=0.3,
+                threshold=0.5,
                 sampling_rate=cls.sample_rate, 
                 # speech_pad_ms=10
-                min_silence_duration_ms = 100,
+                min_silence_duration_ms = 50,
                 # speech_pad_ms = 30,
-                max_speech_duration_s=15
+                max_speech_duration_s=20
                 )
             cls.vac.reset_states()
    
@@ -50,7 +50,7 @@ class VadPipe(BasePipe):
             if start_frame:
                 relative_start_frame = start_frame - self._offset
             if end_frame:
-                relative_end_frame = end_frame - self._offset
+                relative_end_frame = max(0, end_frame - self._offset)
             return relative_start_frame, relative_end_frame
     
     def process(self, in_data: MetaItem) -> MetaItem:
@@ -70,13 +70,10 @@ class VadPipe(BasePipe):
                 self._status = "END" # 音频结束
                 target_audio = source_audio[:rel_end_frame]
                 logging.debug(" 🫷Speech ended, capturing audio up to frame: {}".format(rel_end_frame))
-            elif rel_start_frame is not None and rel_end_frame is not None:
+            else:
                 self._status = 'END'
                 target_audio = source_audio[rel_start_frame:rel_end_frame]
                 logging.debug(" 🔄 Speech segment captured from frame {} to frame {}".format(rel_start_frame, rel_end_frame))
-            else:
-                self._status = 'END'
-                target_audio = np.array([],dtype=np.float32)
                 # logging.debug("❌ No valid speech segment detected, setting status to END")
         else:
             if self._status == 'START':

transcribe/whisper_llm_serve.py

@@ -132,7 +132,7 @@ class WhisperTranscriptionService:
             try:
                 frame_np = self._frame_queue.get(timeout=0.1)
                 frame_np, speech_status = self._apply_voice_activity_detection(frame_np)
-                if frame_np is None:
+                if frame_np is None or len(frame_np) == 0:
                     continue
                 with self.lock:
                     if self.frames_np is None:
@@ -165,19 +165,20 @@ class WhisperTranscriptionService:
         while not self._translate_thread_stop.is_set():
 
             if self.frames_np is None:
-                time.sleep(0.2)
+                time.sleep(0.01)
                 continue
 
-            with self.lock:
-                if len(self.segments_queue) >0:
-                    audio_buffer = self.segments_queue.pop()
-                    partial = False
-                else:
+            
+            if len(self.segments_queue) >0:
+                audio_buffer = self.segments_queue.pop()
+                partial = False
+            else:
+                with self.lock:
                     audio_buffer = self.frames_np[:int(frame_epoch * 1.5 * self.sample_rate)]# 获取 1.5s * epoch 个音频长度
-                    partial = True
+                partial = True
 
             if len(audio_buffer) ==0:
-                time.sleep(0.2)
+                time.sleep(0.01)
                 continue
 
             if len(audio_buffer) < int(self.sample_rate):