Look up UVR (Ultimate Vocal Remover). I believe most of the models predict frequency masks to split out stems.