I'm at an impasse with realtime because it's directly passing audio to the model without a transcription layer, that's why it can understand tone and emotion, but the api is different from the advanced voice mode in app