Agree, but if your use case only works if the flow is as natural as possible with interruptions handled smoothly and minimal latency, openai is currently your best bet.