Researchers at Microsoft have built a text-to-speech AI model called VALL-E that they claim can simulate anyone’s voice using only three seconds of audio.