NVIDIA VoiceChat 11B is an open-weight full-duplex speech model for researchers and voice engineers who want live spoken dialogue with tool calling instead of a stitched ASR-LLM-TTS pipeline. NVIDIA's NemotronLabs team published it on 3 August 2026 as a single network that listens and speaks at the same time.
Core Features
- Unified streaming stack: a Fast Conformer encoder, the Nemotron Nano v2 backbone and an NVIDIA TTS decoder in one 11B model.
- Measured 448 ms smooth turn-taking latency and 480 ms barge-in response on Full-Duplex-Bench 1.0.
- First open full-duplex model with live tool calling, emitted on a separate output channel while speech continues.
- Operator-defined on-hold phrases so the agent keeps talking during slow API calls.
- Hybrid Mamba and Transformer design trained on roughly 550,000 hours of real and synthetic audio.
- Ships as Hugging Face weights plus a signed NIM container on NGC, served with vLLM.
Best For
- Contact-center and automotive teams prototyping barge-in behavior they cannot measure inside a closed API.
- Speech researchers reproducing full-duplex latency numbers on their own H100 or B200 hardware.
- Engineers benchmarking voice agents before committing to a per-minute hosted service.
Pricing
Free to download at $0 under the OpenMDW 1.1 license, with no hosted endpoint and no per-token rate card. The real cost is infrastructure: it needs a single GPU with at least 80 GB of VRAM, such as an A100, H100, RTX 6000 Pro or B200, on x86_64 Linux.
Our Take
The benchmark honesty is the best part. Tool selection reaches 82.5% but argument accuracy sits at 44.2% and pass@1 at 33%, and NVIDIA documents a two-minute context ceiling plus degradation over long conversations. Best for research and architecture study, not shipping. The trade-off is blunt: the checkpoint is labeled research-only, so you cannot launch a product on it. Browse hosted alternatives in AI audio tools on aifreetool.site.
