Merge pull request #807 from alibaba-damo-academy/dev_wjm

jmwang66

2023-08-07 cf8e000a84e888495dcf30c4dbfecea1ee7ab4e2

 funasr/bin/asr_inference_launch.py

@@ -260,8 +260,6 @@
        hotword_list_or_file = None
        clas_scale = 1.0

    if kwargs.get("device", None) == "cpu":
        ngpu = 0
    if ngpu >= 1 and torch.cuda.is_available():
        device = "cuda"
    else:
@@ -565,6 +563,8 @@
        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)

@@ -647,8 +647,7 @@
            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
@@ -1340,7 +1339,7 @@
        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"
@@ -1371,10 +1370,7 @@
        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,