They say it's voice only. I wonder how that works. I worked with full duplex voice llm but the trick is that you feed in intelligence as a third silent participant to the conversation. Would dessa need a second audio stream for that?