Input Normalization

[JunzheJosephZhu]

I'm not sure if my mixing is exactly same as yours, but does your torchaudio read the wav files to int(value is typically around a couple hundred) or to float values between [-1, 1]?

I started with scipy which loads to int, and it caused loss going to NaN at a point. So I switched to librosa which loads to float

[JusperLee]

float values between [-1, 1]. if you started with scipy which loads to int, you will norm audio between [-1,1]