Microsoft's new text-to-speech AI language model, Vall-E, was reportedly found to be capable of mimicking a voice using just a three-second sample recording.