With just a 15-second sample of your voice, OpenAI’s latest model can generate audio of you saying almost anything.