I though about this idea a lot! I really consider to start on it seriously. But for me if we want it to be useful, we need to make it vocal. ML for sound is really harsh though...