NVIDIA's Nemotron VoiceChat Hits 450ms Latency

NVIDIA has put out NemotronLabs VoiceChat 11B, an open speech-to-speech model built to handle live conversation without the usual pipeline of separate speech recognition, language model, and text-to-speech components bolted together. Instead, one 11-billion-parameter network listens and talks at the same time, streaming understanding and generation through a single system rather than passing audio between models and APIs. On Full-Duplex-Bench 1.0, NVIDIA clocks smooth turn-taking latency at 448 milliseconds, and the model can be interrupted mid-sentence: at 480 milliseconds it yields to a user barging in with a take-over rate of 1.00. It's also the first open full-duplex model that keeps a conversation running while calling external tools, routing those calls through a separate output channel so the agent can fill dead air with operator-written lines while an API request finishes.

The weights and container are public and the license is permissive, which puts real deployment questions on the table for teams weighing whether to build on it now. NVIDIA's own documentation, though, lays out where the model breaks down, and that's worth reading closely before anyone plans a rollout.