fix torchaudio load mp3 bug

haoneng.lhn

2023-06-26 e677eb4b13b5388f4351a164a991cea950773a72

 funasr/utils/prepare_data.py

@@ -7,6 +7,7 @@
import numpy as np
import torch.distributed as dist
import torchaudio
import soundfile


def filter_wav_text(data_dir, dataset):
@@ -42,7 +43,11 @@


def wav2num_frame(wav_path, frontend_conf):
    waveform, sampling_rate = torchaudio.load(wav_path)
    try:
        waveform, sampling_rate = torchaudio.load(wav_path)
    except:
        waveform, sampling_rate = soundfile.read(wav_path)
        waveform = np.expand_dims(waveform, axis=0)
    n_frames = (waveform.shape[1] * 1000.0) / (sampling_rate * frontend_conf["frame_shift"] * frontend_conf["lfr_n"])
    feature_dim = frontend_conf["n_mels"] * frontend_conf["lfr_m"]
    return n_frames, feature_dim
@@ -185,7 +190,7 @@
        for i in range(nj):
            path = ""
            for file_name in file_names:
                path = path + os.path.join(split_path, str(i + 1), file_name)
                path = path + " " + os.path.join(split_path, str(i + 1), file_name)
            f_data.write(path + "\n")


@@ -207,10 +212,11 @@
    data_names = args.dataset_conf.get("data_names", "speech,text").split(",")
    data_types = args.dataset_conf.get("data_types", "sound,text").split(",")
    file_names = args.data_file_names.split(",")
    print("data_names: {}, data_types: {}, file_names: {}".format(data_names, data_types, file_names))
    assert len(data_names) == len(data_types) == len(file_names)
    if args.dataset_type == "small":
        args.train_shape_file = [os.path.join(args.data_dir, args.train_set, "{}_shape".format(data_names[0]))]
        args.valid_shape_file = [os.path.join(args.data_dir, args.valid_set, "{}}_shape".format(data_names[0]))]
        args.valid_shape_file = [os.path.join(args.data_dir, args.valid_set, "{}_shape".format(data_names[0]))]
        args.train_data_path_and_name_and_type, args.valid_data_path_and_name_and_type = [], []
        for file_name, data_name, data_type in zip(file_names, data_names, data_types):
            args.train_data_path_and_name_and_type.append(