Would be great to have voice input and voice response! A new LLM type could be added for that. Or download and run with local whisper models.