A compact 4-billion-parameter real-time speech model from Mistral AI for low-latency voice applications.