Nvidia has released NemotronLabs VoiceChat 11B, an open speech-to-speech model built for real-time conversation, according to MarkTechPost, which reported the launch on August 9, 2026.
The headline number is latency. MarkTechPost reports the model handles turn-taking in about 450 milliseconds — 448 ms, by its account. That is roughly the pause you'd expect from a person who is actually listening, rather than the awkward beat of silence most voice assistants leave while they think.
The model is described as "full-duplex," which is the more interesting technical claim. Most voice AI today is half-duplex: you talk, it waits, then it talks, and you wait. Full-duplex means both sides of the conversation can be active at once, so the system can listen while it speaks — the groundwork for being interrupted mid-sentence and adjusting, the way humans do.
It is also speech-to-speech, meaning audio goes in and audio comes out. Conventional voice assistants chain together three separate systems: transcribe the speech to text, run the text through a language model, then synthesize a voice. Every hop adds delay and discards information like tone and emphasis.
MarkTechPost also reports the model supports live tool calling, meaning it can reach out to external functions or services during a conversation rather than only afterward — the difference between an assistant that talks about booking something and one that books it while you're still on the line.
At 11 billion parameters, it is small by frontier standards. Nvidia is releasing it openly.
Why it matters: latency is the main reason talking to AI still feels like talking to a machine, and an open model that closes that gap puts natural-sounding voice interfaces within reach of developers who can't build them from scratch.