add speech start padding 100ms

config.py

@@ -24,7 +24,7 @@ logging.getLogger().addHandler(console_handler)
 # 文字输出长度阈值
 TEXT_THREHOLD = 6
 # 音频段的决策时间
-DESIGN_TIME_THREHOLD = 3
+FRAME_SCOPE_TIME_THREHOLD = 3
 # 最长语音时长
 MAX_SPEECH_DURATION_S = 15
 

transcribe/pipelines/pipe_vad.py

@@ -61,11 +61,10 @@ class VadPipe(BasePipe):
         speech_data  = self._process_speech_chunk(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 is not None and rel_end_frame is None:
                 self._status = "START" # 语音开始
-                target_audio = source_audio[rel_start_frame:]
+                target_audio = source_audio[max(rel_start_frame-100, 0):]
                 logging.debug("🫸 Speech start frame: {}".format(rel_start_frame))
             elif rel_start_frame is None and rel_end_frame is not None:
                 self._status = "END" # 音频结束

transcribe/whisper_llm_serve.py

@@ -153,7 +153,7 @@ class WhisperTranscriptionService:
 
                     elif speech_status == "END" and len(self.frames_np) > 0 and self.frames_np_start_timestamp:
                         time_diff = time.time() - self.frames_np_start_timestamp
-                        if time_diff >= config.DESIGN_TIME_THREHOLD:
+                        if time_diff >= config.FRAME_SCOPE_TIME_THREHOLD:
                             audio_array=self.frames_np.copy()
                             self.full_segments_queue.appendleft(audio_array) # 根据时间是否满足三秒长度 来整合音频块
                             self.frames_np_start_timestamp = None