low latency speech-to-text -> LLM -> text-to-speech