Interesting. How do you handle the latency? Voice AI often feels slow for real-time practice, so I'm curious about your optimization.