A new machine learning system from the team at MIT relies on visual cues to turn mono sound to stereo.