One second voice-to-voice latency with just open models | Not Hacker News!