Haven’t OpenAI models been trained on text + audio for a year or two now?