I don't believe it's a difficult problem - no need for ML/AI since removing vocals was / has been done decades ago already, see karaoke - it's just that nobody's bothered to build it. Plus, adapting a song like that has legal implications; iirc you can't just release a karaoke version of a song for example.
