nemotron-voicechat

The dutchstartup.ai news brief reports that NVIDIA released Nemotron VoiceChat 11B as an open, end-to-end voice-to-voice model for real-time, full-duplex conversations, situating the release alongside Google's separate decision to phase out Google Assistant on Android and Wear OS in favor of Gemini from September 4, 20 The piece frames both developments as evidence that voice-based interaction with AI systems is changing rapidly, noting that the open Nemotron VoiceChat 11B model competes in a space previously dominated by closed, hosted, per-minute voice APIs, though it provides no benchmark detail or deployment specifics for the NVI

AI Weekly reports that NVIDIA released NemotronLabs VoiceChat 11B as the first open-weight full-duplex voice model with live function calling, an end-to-end speech model that skips the conventional ASR → LLM → TTS pipeline in favor of a single hybrid Mamba/Transformer stack. The architecture pairs a Fast Conformer spee The reported model card numbers include roughly 450 ms latency for smooth turn-taking, 480 ms for user interruption, a #2 ranking among open full-duplex models on VoiceBench, and BFCL-v3 tool-calling averages of 56.1% with 82.5% tool selection but only 44.2% argument accuracy and 33% Pass@1. Artificial Analysis benchma

The note.com analysis confirms that on August 3, 2026 NVIDIA released NemotronLabs VoiceChat 11B as an open-weight model that simultaneously handles speech understanding and generation within a single architecture, adopting a full-duplex configuration that allows speaking while listening rather than chaining ASR, LLM, The piece organizes the internal architecture, benchmark capabilities and limitations, and the relationship with the similarly named NIM-hosted Nemotron 3 VoiceChat, including a workflow for trying the model out. It targets technical selectors and implementers of voice AI and conversational agents, framing the release

NVIDIA has opened an Early Access Program for Nemotron 3 VoiceChat, a 12B-parameter, full-duplex speech-to-speech model that unifies ASR, LLM, and TTS into a single architecture for real-time voice agents. The page, hosted on NVIDIA Developer, targets sub-second, natural, interruptible conversations on NVIDIA GPUs and The Early Access Program gives qualified developers hands-on access to the model itself, reference deployment containers, benchmark results, and a guided fine-tuning path for building domain-specific, full-duplex voice agents. Developers can apply to join directly through the NVIDIA Developer portal.