MIT researchers have developed an AI that can mimic human sounds without prior training, using a model of the vocal tract to simulate speech.