NVIDIA Releases VoiceChat 11B for Real-Time Full-Duplex Speech AI

NVIDIA’s NemotronLabs VoiceChat 11B is an open-weight, end-to-end speech-to-speech model designed to replace the conventional pipeline linking automatic speech recognition, a large language model and text-to-speech software. The 11-billion-parameter system is built for full-duplex interaction, allowing an AI agent to listen while speaking, yield when interrupted and respond more naturally — capabilities that could simplify the development of real-time voice assistants.

As of Aug. 10, 2026, NVIDIA has made the VoiceChat 11B weights available for research and testing. The model delivers turn-taking latency of about 450 milliseconds and supports live, dynamic tool calling during a conversation. Running it requires a GPU environment with at least 80GB of VRAM, setting a relatively high deployment threshold even as the open release gives researchers access to its real-time conversational architecture.