Merge pull request #956 from alibaba-damo-academy/chenmengzheAAA-patch-4

chenmengzheAAA

2023-09-14 2a66366be4c2715870e4859fd5a5db6e8a9dc00a

 funasr/bin/asr_inference_launch.py

@@ -236,6 +236,7 @@
        timestamp_infer_config: Union[Path, str] = None,
        timestamp_model_file: Union[Path, str] = None,
        param_dict: dict = None,
        decoding_ind: int = 0,
        **kwargs,
):
    ncpu = kwargs.get("ncpu", 1)
@@ -290,6 +291,7 @@
        nbest=nbest,
        hotword_list_or_file=hotword_list_or_file,
        clas_scale=clas_scale,
        decoding_ind=decoding_ind,
    )

    speech2text = Speech2TextParaformer(**speech2text_kwargs)
@@ -312,6 +314,7 @@
            **kwargs,
    ):

        decoding_ind = None
        hotword_list_or_file = None
        if param_dict is not None:
            hotword_list_or_file = param_dict.get('hotword')
@@ -319,6 +322,8 @@
            hotword_list_or_file = kwargs['hotword']
        if hotword_list_or_file is not None or 'hotword' in kwargs:
            speech2text.hotword_list = speech2text.generate_hotwords_list(hotword_list_or_file)
        if param_dict is not None and "decoding_ind" in param_dict:
            decoding_ind = param_dict["decoding_ind"]

        # 3. Build data-iterator
        if data_path_and_name_and_type is None and raw_inputs is not None:
@@ -365,10 +370,11 @@
            # N-best list of (text, token, token_int, hyp_object)

            time_beg = time.time()
            batch["decoding_ind"] = decoding_ind
            results = speech2text(**batch)
            if len(results) < 1:
                hyp = Hypothesis(score=0.0, scores={}, states={}, yseq=[])
                results = [[" ", ["sil"], [2], hyp, 10, 6]] * nbest
                results = [[" ", ["sil"], [2], hyp, 10, 6, []]] * nbest
            time_end = time.time()
            forward_time = time_end - time_beg
            lfr_factor = results[0][-1]
@@ -409,7 +415,7 @@
                        ibest_writer["rtf"][key] = rtf_cur

                    if text is not None:
                        if use_timestamp and timestamp is not None:
                        if use_timestamp and timestamp is not None and len(timestamp):
                            postprocessed_result = postprocess_utils.sentence_postprocess(token, timestamp)
                        else:
                            postprocessed_result = postprocess_utils.sentence_postprocess(token)
@@ -563,6 +569,8 @@
        if 'hotword' in kwargs:
            hotword_list_or_file = kwargs['hotword']

        speech2vadsegment.vad_model.vad_opts.max_single_segment_time = kwargs.get("max_single_segment_time", 60000)
        batch_size_token_threshold_s = kwargs.get("batch_size_token_threshold_s", int(speech2vadsegment.vad_model.vad_opts.max_single_segment_time*0.67/1000)) * 1000
        batch_size_token = kwargs.get("batch_size_token", 6000)
        print("batch_size_token: ", batch_size_token)

@@ -645,8 +653,7 @@
            beg_idx = 0
            for j, _ in enumerate(range(0, n)):
                batch_size_token_ms_cum += (sorted_data[j][0][1] - sorted_data[j][0][0])
                if j < n - 1 and (batch_size_token_ms_cum + sorted_data[j + 1][0][1] - sorted_data[j + 1][0][
                    0]) < batch_size_token_ms:
                if j < n - 1 and (batch_size_token_ms_cum + sorted_data[j + 1][0][1] - sorted_data[j + 1][0][0]) < batch_size_token_ms and (sorted_data[j + 1][0][1] - sorted_data[j + 1][0][0]) < batch_size_token_threshold_s:
                    continue
                batch_size_token_ms_cum = 0
                end_idx = j + 1
@@ -685,7 +692,7 @@
            text, token, token_int = result[0], result[1], result[2]
            time_stamp = result[4] if len(result[4]) > 0 else None

            if use_timestamp and time_stamp is not None:
            if use_timestamp and time_stamp is not None and len(time_stamp):
                postprocessed_result = postprocess_utils.sentence_postprocess(token, time_stamp)
            else:
                postprocessed_result = postprocess_utils.sentence_postprocess(token)
@@ -1288,7 +1295,8 @@
        quantize_modules: Optional[List[str]] = None,
        quantize_dtype: Optional[str] = "float16",
        streaming: Optional[bool] = False,
        simu_streaming: Optional[bool] = False,
        fake_streaming: Optional[bool] = False,
        full_utt: Optional[bool] = False,
        chunk_size: Optional[int] = 16,
        left_context: Optional[int] = 16,
        right_context: Optional[int] = 0,
@@ -1338,7 +1346,7 @@
        format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s",
    )

    if ngpu >= 1:
    if ngpu >= 1 and torch.cuda.is_available():
        device = "cuda"
    else:
        device = "cpu"
@@ -1364,15 +1372,13 @@
        quantize_modules=quantize_modules,
        quantize_dtype=quantize_dtype,
        streaming=streaming,
        simu_streaming=simu_streaming,
        fake_streaming=fake_streaming,
        full_utt=full_utt,
        chunk_size=chunk_size,
        left_context=left_context,
        right_context=right_context,
    )
    speech2text = Speech2TextTransducer.from_pretrained(
        model_tag=model_tag,
        **speech2text_kwargs,
    )
    speech2text = Speech2TextTransducer(**speech2text_kwargs)

    def _forward(data_path_and_name_and_type,
                 raw_inputs: Union[np.ndarray, torch.Tensor] = None,
@@ -1418,14 +1424,16 @@
                        _end = (i + 1) * speech2text._ctx

                        speech2text.streaming_decode(
                            speech[i * speech2text._ctx: _end], is_final=False
                            speech[i * speech2text._ctx: _end + speech2text._right_ctx], is_final=False
                        )

                    final_hyps = speech2text.streaming_decode(
                        speech[_end: len(speech)], is_final=True
                    )
                elif speech2text.simu_streaming:
                    final_hyps = speech2text.simu_streaming_decode(**batch)
                elif speech2text.fake_streaming:
                    final_hyps = speech2text.fake_streaming_decode(**batch)
                elif speech2text.full_utt:
                    final_hyps = speech2text.full_utt_decode(**batch)
                else:
                    final_hyps = speech2text(**batch)

@@ -1813,7 +1821,8 @@
    group.add_argument("--lm_weight", type=float, default=1.0, help="RNNLM weight")
    group.add_argument("--ngram_weight", type=float, default=0.9, help="ngram weight")
    group.add_argument("--streaming", type=str2bool, default=False)
    group.add_argument("--simu_streaming", type=str2bool, default=False)
    group.add_argument("--fake_streaming", type=str2bool, default=False)
    group.add_argument("--full_utt", type=str2bool, default=False)
    group.add_argument("--chunk_size", type=int, default=16)
    group.add_argument("--left_context", type=int, default=16)
    group.add_argument("--right_context", type=int, default=0)