>>21969931(caught me right before going to sleep, lucky you)
for songs that have an official instrumental version it's super easy, just take both the voiced and voiceless songs, overlap them, and invert the waveform of either one of them. you can do this very easily with audacity. usually you want the hq versions because lossy formats like mp3 may introduce a few artifacts that will make the waveforms not completely cancel each other out.
other than that it gets very tricky since you need to devise some sort of algorithm that can decently recognize voice so it can isolate it, or you could more crudely start filtering out the frequencies that don't correspond to an average human voice (though i doubt ANYONE does it that way). some websites and software are actually surprisingly decent at that so i may give 3:12 a shot tomorrow. (i wanna learn how to do this procedure myself one day though)