[Discussion] Why is it useful to train just masks of stft?

[justStarG]

Spleeter trains many masks of stft to split songs. But why does it work? Is is possible to get a better model if I just input stft feature and output stft of each instrument?

[ashbeats]

If you're referring to this fn: https://github.com/deezer/spleeter/blob/9629dcca9f846efbd32699d40201b97017211957/spleeter/utils/tensor.py#L117

Zero-padding / Masking just helps make the plot easier to visualize.

It does not add extra information in the F or T domain. Instead, there are more F bins, in tighter clusters in the F domain.

and it creates a cleaner Sinc interpolation.