Merge pull request #932 from alibaba-damo-academy/dev_lhn

hnluo

2023-09-11 9fcb3cc06b4e324f0913d2f61b89becc2baeef1b

 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:
@@ -365,6 +370,7 @@
            # 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=[])
@@ -1290,6 +1296,7 @@
        quantize_dtype: Optional[str] = "float16",
        streaming: Optional[bool] = False,
        simu_streaming: Optional[bool] = False,
        full_utt: Optional[bool] = False,
        chunk_size: Optional[int] = 16,
        left_context: Optional[int] = 16,
        right_context: Optional[int] = 0,
@@ -1366,6 +1373,7 @@
        quantize_dtype=quantize_dtype,
        streaming=streaming,
        simu_streaming=simu_streaming,
        full_utt=full_utt,
        chunk_size=chunk_size,
        left_context=left_context,
        right_context=right_context,
@@ -1416,7 +1424,7 @@
                        _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(
@@ -1424,6 +1432,8 @@
                    )
                elif speech2text.simu_streaming:
                    final_hyps = speech2text.simu_streaming_decode(**batch)
                elif speech2text.full_utt:
                    final_hyps = speech2text.full_utt_decode(**batch)
                else:
                    final_hyps = speech2text(**batch)

@@ -1782,6 +1792,12 @@
        default=1,
        help="The batch size for inference",
    )
    group.add_argument(
        "--decoding_ind",
        type=int,
        default=0,
        help="chunk select for chunk encoder",
    )
    group.add_argument("--nbest", type=int, default=5, help="Output N-best hypotheses")
    group.add_argument("--beam_size", type=int, default=20, help="Beam size")
    group.add_argument("--penalty", type=float, default=0.0, help="Insertion penalty")
@@ -1812,6 +1828,7 @@
    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("--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)