An AI text-to-speech model called VALL-E can copy a speaker’s tone and environment and have them say anything, starting with only a 3-second long sample.