Show HN: Caramel: a speech-to-speech conversation model trained only on audio