Update e2e_asr_mfcca.py

yufan

2023-04-14 d268a4360f29c3f3160fd762e43ffaa86a15da5d

 funasr/runtime/python/onnxruntime/funasr_onnx/vad_bin.py

@@ -53,39 +53,56 @@
      proj_dim = self.encoder_conf["proj_dim"]
      lorder = self.encoder_conf["lorder"]
      for i in range(fsmn_layers):
         cache = np.zeros(1, proj_dim, lorder-1, 1).astype(np.float32)
         cache = np.zeros((1, proj_dim, lorder-1, 1)).astype(np.float32)
         in_cache.append(cache)
      return in_cache
      
   
   def __call__(self, wav_content: Union[str, np.ndarray, List[str]], **kwargs) -> List:
      waveform_list = self.load_data(wav_content, self.frontend.opts.frame_opts.samp_freq)
   def __call__(self, audio_in: Union[str, np.ndarray, List[str]], **kwargs) -> List:
      waveform_list = self.load_data(audio_in, self.frontend.opts.frame_opts.samp_freq)
      waveform_nums = len(waveform_list)
      is_final = kwargs.get('kwargs', False)

      asr_res = []
      segments = [[]] * self.batch_size
      for beg_idx in range(0, waveform_nums, self.batch_size):
         
         end_idx = min(waveform_nums, beg_idx + self.batch_size)
         waveform = waveform_list[beg_idx:end_idx]
         feats, feats_len = self.extract_feat(waveform)
         waveform = np.array(waveform)
         param_dict = kwargs.get('param_dict', dict())
         in_cache = param_dict.get('cache', list())
         in_cache = param_dict.get('in_cache', list())
         in_cache = self.prepare_cache(in_cache)
         try:
            inputs = [feats]
            inputs.extend(in_cache)
            scores, out_caches = self.infer(inputs)
            param_dict['cache'] = out_caches
            segments = self.vad_scorer(scores, waveform[0][None, :], is_final=is_final, max_end_sil=self.max_end_sil)
            t_offset = 0
            step = int(min(feats_len.max(), 6000))
            for t_offset in range(0, int(feats_len), min(step, feats_len - t_offset)):
               if t_offset + step >= feats_len - 1:
                  step = feats_len - t_offset
                  is_final = True
               else:
                  is_final = False
               feats_package = feats[:, t_offset:int(t_offset + step), :]
               waveform_package = waveform[:, t_offset * 160:min(waveform.shape[-1], (int(t_offset + step) - 1) * 160 + 400)]

               inputs = [feats_package]
               # inputs = [feats]
               inputs.extend(in_cache)
               scores, out_caches = self.infer(inputs)
               in_cache = out_caches
               segments_part = self.vad_scorer(scores, waveform_package, is_final=is_final, max_end_sil=self.max_end_sil, online=False)
               # segments = self.vad_scorer(scores, waveform[0][None, :], is_final=is_final, max_end_sil=self.max_end_sil)

               if segments_part:
                  for batch_num in range(0, self.batch_size):
                     segments[batch_num] += segments_part[batch_num]
            
         except ONNXRuntimeError:
            # logging.warning(traceback.format_exc())
            logging.warning("input wav is silence or noise")
            segments = ''
         asr_res.append(segments)
   
      return asr_res
      return segments

   def load_data(self,
                 wav_content: Union[str, np.ndarray, List[str]], fs: int = None) -> List:
@@ -134,4 +151,4 @@
      outputs = self.ort_infer(feats)
      scores, out_caches = outputs[0], outputs[1:]
      return scores, out_caches