This demo feels a lot like GPT-V. Like they've gotten a lot of the latencies down, but it's doing the same thing GPT was doing previously with transcription after silence detection and TTS of the output.