update with main (#1783)

zhifu gao

2024-06-04 3b0526e7be3565c42007313b90a018a2f8c8dff1

 runtime/python/onnxruntime/funasr_onnx/paraformer_bin.py

@@ -326,6 +326,9 @@
    def __call__(
        self, wav_content: Union[str, np.ndarray, List[str]], hotwords: str, **kwargs
    ) -> List:
    # def __call__(
    #     self, waveform_list:list, hotwords: str, **kwargs
    # ) -> List:
        # make hotword list
        hotwords, hotwords_length = self.proc_hotword(hotwords)
        # import pdb; pdb.set_trace()
@@ -345,15 +348,47 @@
            try:
                outputs = self.bb_infer(feats, feats_len, bias_embed)
                am_scores, valid_token_lens = outputs[0], outputs[1]

                if len(outputs) == 4:
                    # for BiCifParaformer Inference
                    us_alphas, us_peaks = outputs[2], outputs[3]
                else:
                    us_alphas, us_peaks = None, None

            except ONNXRuntimeError:
                # logging.warning(traceback.format_exc())
                logging.warning("input wav is silence or noise")
                preds = [""]
            else:
                preds = self.decode(am_scores, valid_token_lens)
                for pred in preds:
                    pred = sentence_postprocess(pred)
                    asr_res.append({"preds": pred})
                if us_peaks is None:
                    for pred in preds:
                        if self.language == "en-bpe":
                            pred = sentence_postprocess_sentencepiece(pred)
                        else:
                            pred = sentence_postprocess(pred)
                        asr_res.append({"preds": pred})
                else:
                    for pred, us_peaks_ in zip(preds, us_peaks):
                        raw_tokens = pred
                        timestamp, timestamp_raw = time_stamp_lfr6_onnx(
                            us_peaks_, copy.copy(raw_tokens)
                        )
                        text_proc, timestamp_proc, _ = sentence_postprocess(
                            raw_tokens, timestamp_raw
                        )
                        # logging.warning(timestamp)
                        if len(self.plot_timestamp_to):
                            self.plot_wave_timestamp(
                                waveform_list[0], timestamp, self.plot_timestamp_to
                            )
                        asr_res.append(
                            {
                                "preds": text_proc,
                                "timestamp": timestamp_proc,
                                "raw_tokens": raw_tokens,
                            }
                        )
        return asr_res

    def proc_hotword(self, hotwords):