vad parameters v1 test

transcribe/pipelines/pipe_vad.py

@@ -29,7 +29,7 @@ def collect_chunks(tss: List[dict], wav: torch.Tensor, sample_rate: int = 16000)
 def collect_chunks_improved(tss: List[dict], wav: torch.Tensor, sample_rate: int = 16000):
     chunks = []
     silent_samples = int(0.3 * sample_rate)  # 300ms 的静音样本数
-    silence = torch.zeros(silent_samples)  # 创建300ms的静音
+    # silence = torch.zeros(silent_samples)  # 创建300ms的静音
     min_gap_samples = int(0.1 * sample_rate)  # 最小间隔阈值 (100ms)
     
     # 对时间戳进行简单的平滑处理
@@ -74,22 +74,22 @@ class VadPipe(BasePipe):
     def init(cls):
         if cls.model is None:
             cls.model = SileroVADProcessor(
-            activate_threshold=0.45,       # 降低以捕获更多音频
+            activate_threshold=0.3,       # 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
+            min_silence_duration=500,     # 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 = FixedVADIterator(
+                cls.model.silero_vad, 
+                sampling_rate=cls.sample_rate, 
+                threshold=0.3,
+                speech_pad_ms=10
+                )
             cls.vac.reset_states()
    
-    
-    def get_previous_buffer(self):
-        if len(self.previous_buffer) == 2:
-            return self.previous_buffer[-1]
-        return np.array([], dtype=np.float32)
-    
+
 
     # def reduce_noise(self, data):
     #     return nr.reduce_noise(y=data, sr=self.sample_rate)
@@ -109,10 +109,10 @@ class VadPipe(BasePipe):
     def process(self, in_data: MetaItem) -> MetaItem:
         if self._offset == 0:
             self.vac.reset_states()
-
+        silence_audio_100ms = np.zeros(int(0.1*self.sample_rate))
         source_audio = np.frombuffer(in_data.source_audio, dtype=np.float32)
         speech_data  = self._process_speech_chunk(source_audio)
-        self._offset += len(source_audio)
+       
         if speech_data: # 表示有音频的变化点出现
             rel_start_frame, rel_end_frame = speech_data 
             if rel_start_frame and not rel_end_frame:
@@ -130,16 +130,17 @@ class VadPipe(BasePipe):
             else:
                 self._status = 'END'
                 target_audio = np.array([],dtype=np.float32)
-                logging.debug("❌ No valid speech segment detected, setting status to END")
+                # logging.debug("❌ No valid speech segment detected, setting status to END")
         else:
             if self._status == 'START':
                 target_audio = source_audio
-                logging.debug("🔊 Continuing to capture audio as speech is still ongoing")
+                # logging.debug("🔊 Continuing to capture audio as speech is still ongoing")
             else: # end
                 target_audio = np.array([],dtype=np.float32)
-                self._status = 'END'
-                logging.debug("❌ No speech detected, setting status to END")
-       
+                # self._status = 'END'
+                # logging.debug("❌ No speech detected, setting status to END")
+
+        self._offset += len(source_audio)
 
         in_data.audio = target_audio.tobytes()
         in_data.source_audio = b''

transcribe/strategy.py

@@ -111,7 +111,7 @@ class TranscriptChunk:
             return 0
 
         score =  self._calculate_similarity(self.join(), chunk.join())
-        logger.debug(f"Compare: {self.join()} vs {chunk.join()} : {score}")
+        # logger.debug(f"Compare: {self.join()} vs {chunk.join()} : {score}")
         return score
 
     def only_punctuation(self)->bool:

transcribe/whisper_llm_serve.py

@@ -159,15 +159,15 @@ class WhisperTranscriptionService:
         frames = frame_np.copy()
 
         # 音频过短时的处理
-        if len(frames) <= 10:
+        if len(frames) < self.sample_rate:
             # 极短音频段，清空并返回None
             # self._update_audio_buffer(len(frames))
             return None
-        if len(frames) < self.sample_rate:
-            # 不足一秒的音频，补充静音
-            silence_audio = np.zeros((self.sample_rate + 1000,), dtype=np.float32)
-            silence_audio[-len(frames):] = frames
-            return silence_audio.copy()
+        # if len(frames) < self.sample_rate:
+        #     # 不足一秒的音频，补充静音
+        #     silence_audio = np.zeros((self.sample_rate + 1000,), dtype=np.float32)
+        #     silence_audio[-len(frames):] = frames
+        #     return silence_audio.copy()
 
         return frames