fix error in get audio chunk

transcribe/pipelines/pipe_vad.py

@@ -41,6 +41,7 @@ class VadPipe(BasePipe):
     def process(self, in_data: MetaItem) -> MetaItem:
         source_audio = in_data.source_audio
         source_audio = np.frombuffer(source_audio, dtype=np.float32)
+        
         send_audio = b""
         speech_timestamps = get_speech_timestamps(torch.Tensor(source_audio), self.model.silero_vad, sampling_rate=16000)
         

transcribe/whisper_llm_serve.py

@@ -152,13 +152,14 @@ class PyWhiperCppServe(ServeClientBase):
                 logger.info("Exiting speech to text thread")
                 break
     
-            if self.frames_np is None:
+            if self.frames_np is None :
                 time.sleep(0.02)  # wait for any audio to arrive
                 continue
             
             audio_buffer = self.get_audio_chunk_for_processing()
-            if audio_buffer.shape[0] < self.sample_rate * 2:
-                time.sleep(0.02)
+            audio_duration = (len(audio_buffer) / self.sample_rate)
+            if audio_duration<0.5 :
+                time.sleep(0.02)  # wait for any audio to arrive
                 continue
             # c+= 1
             # name = f"dev-{c}.wav"
@@ -172,6 +173,8 @@ class PyWhiperCppServe(ServeClientBase):
             #     break
             # except Exception as e:
             #     logger.error(f"{e}")
+            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)
@@ -239,8 +242,15 @@ class PyWhiperCppServe(ServeClientBase):
 
     def get_audio_chunk_for_processing(self):
         self.vad_merge()
-
-        return self.frames_np.copy()
+        # silence_audio = np.zeros((self.sample_rate+100,), 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)
 
     def stop(self):
         self._translate_thread_stop.set()