update

shixian.shi

2023-06-27 b0f4910de6dc91c13828026fb5bdd4f15d8636f3

 funasr/bin/asr_inference_launch.py

@@ -19,6 +19,7 @@
import numpy as np
import torch
import torchaudio
import soundfile
import yaml
from typeguard import check_argument_types

@@ -259,6 +260,7 @@
        export_mode = param_dict.get("export_mode", False)
    else:
        hotword_list_or_file = None
    clas_scale = param_dict.get('clas_scale', 1.0)

    if kwargs.get("device", None) == "cpu":
        ngpu = 0
@@ -291,6 +293,7 @@
        penalty=penalty,
        nbest=nbest,
        hotword_list_or_file=hotword_list_or_file,
        clas_scale=clas_scale,
    )

    speech2text = Speech2TextParaformer(**speech2text_kwargs)
@@ -863,7 +866,13 @@
            raw_inputs = _load_bytes(data_path_and_name_and_type[0])
            raw_inputs = torch.tensor(raw_inputs)
        if data_path_and_name_and_type is not None and data_path_and_name_and_type[2] == "sound":
            raw_inputs = torchaudio.load(data_path_and_name_and_type[0])[0][0]
            try:
                raw_inputs = torchaudio.load(data_path_and_name_and_type[0])[0][0]
            except:
                raw_inputs = soundfile.read(data_path_and_name_and_type[0], dtype='float32')[0]
                if raw_inputs.ndim == 2:
                    raw_inputs = raw_inputs[:, 0]
                raw_inputs = torch.tensor(raw_inputs)
        if data_path_and_name_and_type is None and raw_inputs is not None:
            if isinstance(raw_inputs, np.ndarray):
                raw_inputs = torch.tensor(raw_inputs)