Originally Posted by
Anjok
If you mean cleaner in that there's less instrumental bleed through, yes much cleaner. However, if you mean clean as in having the same overall sound quality as the generated instrumental, then no. Not yet. The issue isn't the AI either. It's because I don't have enough official acapellas to use in a dataset.
The way to train this AI to extract vocals with retained quality would be to add the official acapellas to the "instruments" folder and the mixes to the "mixture" folder. I actually coded a version of the GUI as "Ultimate Vocal Extractor". I just need a bigger dataset to train with, so it's still a work in progress.