The deep-learning machine can match digital sounds to visuals in a way that fools people most of the time.