Sampling has become an incredibly important technique in modern music, relied upon by artists in genres as diverse as electronic music, hip hop, indy rock, and even contemporary classical. However, the vast majority of music isn't released in stem format (separate tracks for vocals, drums, bass, guitar, synth, etc.), which can make sampling or remixing most music very difficult to impossible. What if there was software that could automatically "extract" stems from an un-stemmed song? How many exciting possibilities would this open up for music?
It may not sound like it, but this is actually an extraordinarily difficult computing challenge to solve. It's a bit like trying to extract the eggs, flour, and sugar from cookie dough. Realistically, how might this be accomplished?
One possibility would be to use stemmed and un-stemmed versions of the same songs as training data for machine learning algorithms. With enough data, these algorithms could hopefully be able to learn to classify, or "lock on", to various musical elements that comprise a track. Using the same techniques, the various methods that producers use to manually isolate musical elements could perhaps be automated. And finally, if a particular note of a musical element can't be completely isolated from the "noise" of the rest of the track, perhaps that note could be "simulated", or recreated, similar to a MIDI note.
(Note: You will only be charged
if this prize is won)
All images courtesy of Wikipedia.