update tp_inference outputs

shixian.shi

2023-03-09 900346d13d35552fad1b4bd1151478d875a6fbf8

 funasr/bin/tp_inference.py

@@ -112,6 +112,9 @@
        tp_model, tp_train_args = ASRTask.build_model_from_file(
            timestamp_infer_config, timestamp_model_file, device
        )
        if 'cuda' in device:
            tp_model = tp_model.cuda()

        frontend = None
        if tp_train_args.frontend is not None:
            frontend = WavFrontend(cmvn_file=timestamp_cmvn_file, **tp_train_args.frontend_conf)
@@ -148,11 +151,11 @@
        # Input as audio signal
        if isinstance(speech, np.ndarray):
            speech = torch.tensor(speech)

        if self.frontend is not None:
            feats, feats_len = self.frontend.forward(speech, speech_lengths)
            feats = to_device(feats, device=self.device)
            feats_len = feats_len.int()
            self.tp_model.frontend = None
        else:
            feats = speech
            feats_len = speech_lengths
@@ -240,7 +243,6 @@
        device = "cuda"
    else:
        device = "cpu"

    # 1. Set random-seed
    set_all_random_seed(seed)

@@ -302,7 +304,9 @@
                token = speechtext2timestamp.converter.ids2tokens(batch['text'][batch_id])
                ts_str, ts_list = time_stamp_lfr6_advance(us_alphas[batch_id], us_cif_peak[batch_id], token)
                logging.warning(ts_str)
                tp_result_list.append({'text':"".join([i for i in token if i != '<sil>']), 'timestamp': ts_list})
                item = {'key': key, 'value': ts_str, 'timestamp':ts_list}
                # tp_result_list.append({'text':"".join([i for i in token if i != '<sil>']), 'timestamp': ts_list})
                tp_result_list.append(item)
        return tp_result_list

    return _forward