NVIDIA Releases NemotronLabs-VoiceChat-11B for Full-Duplex Voice Interaction
NVIDIA has released NVIDIA-NemotronLabs-VoiceChat-11B on Hugging Face, an 11-billion-parameter model designed for full-duplex voice interaction. This model allows for more natural, real-time conversational voice AI experiences. This release represents a significant advancement for local multimodal AI, enabling developers to build voice assistants that can listen and speak simultaneously. It moves voice AI away from laggy, walkie-talkie-style interactions toward natural, fluid human-like conversations. The model is hosted on Hugging Face as an open-weights release, making it accessible for local deployment and customization. It is designed to handle conversational dynamics like interruptions and real-time turn-taking.
## BACKGROUND
Traditional voice AI systems operate in half-duplex mode, where the user and the AI must take turns speaking, similar to using a walkie-talkie. Full-duplex voice interaction allows the AI to listen and process audio while simultaneously speaking, enabling it to handle interruptions and backchannel cues naturally.