fix torchaudio load mp3 bug

haoneng.lhn

2023-06-26 e677eb4b13b5388f4351a164a991cea950773a72

 funasr/utils/wav_utils.py

@@ -11,6 +11,7 @@
import numpy as np
import torch
import torchaudio
import soundfile
import torchaudio.compliance.kaldi as kaldi


@@ -162,7 +163,11 @@
        waveform = torch.from_numpy(waveform.reshape(1, -1))
    else:
        # load pcm from wav, and resample
        waveform, audio_sr = torchaudio.load(wav_file)
        try:
            waveform, audio_sr = torchaudio.load(wav_file)
        except:
            waveform, audio_sr = soundfile.read(wav_file)
            waveform = torch.tensor(np.expand_dims(waveform, axis=0))
        waveform = waveform * (1 << 15)
        waveform = torch_resample(waveform, audio_sr, model_sr)

@@ -181,7 +186,11 @@


def wav2num_frame(wav_path, frontend_conf):
    waveform, sampling_rate = torchaudio.load(wav_path)
    try:
        waveform, audio_sr = torchaudio.load(wav_file)
    except:
        waveform, audio_sr = soundfile.read(wav_file)
        waveform = torch.tensor(np.expand_dims(waveform, axis=0))
    speech_length = (waveform.shape[1] / sampling_rate) * 1000.
    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"]