NVIDIA published the technical report for NemotronLabs VoiceChat on 18 September, an open-weights full-duplex speech-to-speech model that can also call external tools. The 11-billion-parameter weights went up on Hugging Face earlier this month; the paper is what now documents how the system is built and how it scores.
Full-duplex means the model listens and speaks at the same time, so it can be interrupted mid-sentence and can cope with the noises people make while someone else is talking. The architecture puts a streaming speech encoder, a decoder-only language model, an RNN-T branch for transcription and a streaming speech synthesiser into one pipeline, with a parallel output stream reserved for function calls.
On the Full-Duplex-Bench suite the authors report the lowest pause-handling takeover rates among the open-weight systems they tested, a response-quality score of 4.33 out of 5 after an interruption, and resumption of the response after a listener’s murmured acknowledgement in 93% of cases. Tool use is the weaker half: 82.5% F1 on choosing the right tool, with the report itself calling argument accuracy and end-to-end execution areas that need work.
Every figure here is the authors’ own, on a benchmark they did not create but did select.