this is contrasted to how openai models work in vapi where the tts is initiated soon after the first tokens are received