The current implementation contains a padding issue with librosa.stft.
As a result, the output of torch.stft and librosa.stft is different, which leads to accuracy degradation.
In this PR, I fixed that padding mode from 'constant,' which is the default HP of librosa.stft, into 'reflect,' which is the default HP of torch.stft.
The current implementation contains a padding issue with librosa.stft. As a result, the output of torch.stft and librosa.stft is different, which leads to accuracy degradation. In this PR, I fixed that padding mode from 'constant,' which is the default HP of librosa.stft, into 'reflect,' which is the default HP of torch.stft.