i'm guessing the model is doing something similar, and has a fairly confident set of the possible phrases to map sounds to.