NVIDIA Open-Sources VoiceChat, a Full-Duplex Speech-to-Speech Model with Tool Calling

chaumian · x · 2026-09-22

NVIDIA's NemotronLabs team released a paper on VoiceChat, an open full-duplex speech-to-speech model that stands out for supporting tool calling during real-time voice conversations.

Unlike turn-based voice assistants, the full-duplex design lets the model listen and speak simultaneously, closer to natural human conversation. Tool calling enables the model to trigger external actions within a purely voice-native interaction—a key step toward practical voice agents. The paper lists a large author team of NVIDIA speech researchers.

Related event: NVIDIA Open-Sources Full-Duplex VoiceChat with Tool Calling(2 posts)→

Original post →

More from Multimodal

Multimodal channel →