update vad config parameter

transcribe/pipelines/pipe_vad.py

@@ -17,12 +17,12 @@ class VadPipe(BasePipe):
     def init(cls):
         if cls.model is None:
             cls.model = SileroVADProcessor(
-            activate_threshold=0.5,
-            fusion_threshold=0.3,
-            min_speech_duration=0.25,
-            max_speech_duration=20,
-            min_silence_duration=250,
-            sample_rate=cls.sample_rate
+            activate_threshold=0.4,       # 降低以捕获更多音频
+            fusion_threshold=0.45,        # 提高以更好地融合语音片段
+            min_speech_duration=0.2,      # 略微降低以捕获短音节
+            max_speech_duration=20,       # 保持不变
+            min_silence_duration=300,     # 增加到300毫秒，允许说话间的自然停顿
+            sample_rate=cls.sample_rate # 采样率，音频信号的采样频率
         )
             cls.vac = FixedVADIterator(cls.model.silero_vad, sampling_rate=cls.sample_rate,)
             cls.vac.reset_states()

transcribe/whisper_llm_serve.py

@@ -146,7 +146,7 @@ class PyWhiperCppServe(ServeClientBase):
         return None, left_watch_string, right_watch_string, is_end_sentence
 
     def speech_to_text(self):
-        # c = 0
+        c = 0
         while not self._translate_thread_stop.is_set():
             if self.exit:
                 logger.info("Exiting speech to text thread")
@@ -154,13 +154,14 @@ class PyWhiperCppServe(ServeClientBase):
     
             if self.frames_np is None :
                 time.sleep(0.02)  # wait for any audio to arrive
+                logger.info("waiting for client data...")
                 continue
-            
+
             audio_buffer = self.get_audio_chunk_for_processing()
-            audio_duration = (len(audio_buffer) / self.sample_rate)
-            if audio_duration<0.5 :
+            if audio_buffer is None:
                 time.sleep(0.02)  # wait for any audio to arrive
                 continue
+            
             # c+= 1
             # name = f"dev-{c}.wav"
             # save_to_wave(name, audio_buffer)
@@ -173,8 +174,8 @@ class PyWhiperCppServe(ServeClientBase):
             #     break
             # except Exception as e:
             #     logger.error(f"{e}")
-            if (time_delay := (1 - audio_duration)) > 0:
-                time.sleep(time_delay)
+            # if (time_delay := (1 - audio_duration)) > 0:
+            #     time.sleep(time_delay)
     
     def handle_transcription_output(self, segments, audio_buffer):
         texts  = self.text_sep.join(i.text for i in segments)
@@ -242,15 +243,18 @@ class PyWhiperCppServe(ServeClientBase):
 
     def get_audio_chunk_for_processing(self):
         self.vad_merge()
-        # silence_audio = np.zeros((self.sample_rate+100,), dtype=np.float32)
+        silence_audio = np.zeros((self.sample_rate+1000,), dtype=np.float32)
         frames = self.frames_np.copy()
-        #todo 如果补空白音频就会导致 幻听
-        # if 0< len(frames) < self.sample_rate:
-        #     silence_audio[-len(frames):] = frames
-        #     return silence_audio.copy()
-        if len(frames) > self.sample_rate:
-            return frames.copy()
-        return np.zeros((0,), dtype=np.float32)
+            # 添加对非常短音频的处理
+        if len(frames) <= 100:
+            # 对于极短的音频段(<=100帧)，直接返回空音频
+            self.update_audio_buffer(len(frames))
+            return None
+        elif len(frames) < self.sample_rate:
+            silence_audio[-len(frames):] = frames
+            return silence_audio.copy()
+        return frames.copy()
+
 
     def stop(self):
         self._translate_thread_stop.set()