How use the torchaudio.tansforms.Melspectrogram to replace the mel_spectrogram function in meldataset.py?

[liyihao17]

I tried to replace the mel_spectrogram function by torchaudio.tansforms.Melspectrogram, but the output of these two function is difference, which lead to the failure of transforming the log mel spectrogram to wave file.

[liyihao17]

I set the mel_spectorgram function like :
mel = mel_spectrogram(wav, n_fft=1024, num_mels=80, sampling_rate=16000, hop_size=256, win_size=1024, fmin=0, fmax=8000)

and set the torchaudio class like :
torchaudio.transforms.MelSpectrogram(sample_rate=16000, n_fft=1024, hop_length=256, f_min=0, f_max=8000, n_mels=80, win_length=1024, window_fn=torch.hann_window, center=False, onesided=True)
mel = torch.log(torch.clamp(mel, min=1e-10))

this two function cannot output the same mel spectorgram

[sophiefy]

Seems that the mel filter banks in librosa and torchaudio are a bit different, resulting in different mel-spectrogram.