Hacker News Reader

I haven't tried this idea, yet, however: it should be possible to use mixdown -> stem AI/ML models to completely remove the "recognizable" part of the song to an algorithm, leaving only the "vocals" portion.

I know experimentally youtube nearly immediately deletes/unlists a few artists' songs, which i suppose one could use to test the theory.

View on HN

Definitely possible; for example: https://github.com/deezer/spleeter

Yes, this (spleeter) is specifically what i was thinking of. I have to try it again, the version i have is a couple years old and there is a lot of artifacting. I also happen to have a fairly large library of stems, which i could probably use to train the model some more.

Reply on news.ycombinator.com