New text-to-speech research shows AI can “unlearn” how to mimic specific voices, reducing misuse risks without losing overall voice performance.