Upload a reference recording and its exact transcript, then synthesize new speech with the same voice.