fix vad bug

transcribe/pipelines/pipe_vad.py

@@ -42,10 +42,12 @@ class VadPipe(BasePipe):
         source_audio = in_data.source_audio
         source_audio = np.frombuffer(source_audio, dtype=np.float32)
         send_audio = b""
-        speech_timestamps = get_speech_timestamps(source_audio, self.model.silero_vad, sampling_rate=16000)
+        speech_timestamps = get_speech_timestamps(torch.Tensor(source_audio), self.model.silero_vad, sampling_rate=16000)
+        
         if speech_timestamps:
             send_audio = collect_chunks(speech_timestamps, torch.Tensor(source_audio))
             send_audio = send_audio.numpy()
+            in_data.audio = send_audio
             # send_audio = self.reduce_noise(send_audio).tobytes()
         in_data.source_audio = b""
         return in_data

transcribe/whisper_llm_serve.py

@@ -68,9 +68,8 @@ class PyWhiperCppServe(ServeClientBase):
         with self.lock:
             frame = self.frames_np.copy()
             item = self._translate_pipes.voice_detect(frame.tobytes())
-            if item.audio != b'':
-                frame_np = np.frombuffer(item.audio, dtype=np.float32)
-                self.frames_np = frame_np.copy()
+            frame_np = np.frombuffer(item.audio, dtype=np.float32)
+            self.frames_np = frame_np.copy()
      
     
     def get_frame_from_queue(self,):
@@ -103,6 +102,7 @@ class PyWhiperCppServe(ServeClientBase):
 
         item = self._translate_pipes.transcrible(audio_buffer.tobytes(), self.language)
         segments = item.segments
+        log_block("Whisper transcrible out", f"{''.join(seg.text for seg in segments)}", "")
         log_block("Whisper transcrible time", f"{(time.perf_counter() - start_time):.3f}", "s")
        
         return segments
@@ -184,6 +184,7 @@ class PyWhiperCppServe(ServeClientBase):
         if last_cut_index:
             self.update_audio_buffer(last_cut_index)
             # 句子或者短句的提交
+            log_block("Whisper string lock ", f"{left_string}",)
             self._segment_manager.handle(left_string).commit(is_end_sentence)
             self._segment_manager.handle(right_string)