Nemotron 3 Diarization provides a way for developers to maintain clean transcripts by distinguishing between multiple talkers in a single audio stream. The 100M parameter model, now available as open weights on Hugging Face, can identify up to 8 speakers and track identities even when voices overlap. Nvidia released the tool under a commercial friendly license to allow for deployment on local devices or in the cloud.
The tool is designed for use in voice agents and real time AI applications to improve how machines recognize who is speaking to them. It features day zero integration with Transformers and can pair with existing transcription models. Early testing involving a DGX Spark and a Reachy Mini robot demonstrated the system's ability to identify a new voice, request a name, and recall it.
Key sources
- SOURCE@nvidiaai“Our new Nemotron 3 Diarization model tracks who spoke when, even when voices overlap”x.com
- SUPPORT@andimarafioti“NVIDIA is open-sourcing Nemotron 3 Diarization: a model that can reliably track speakers in live conversations, under a commercial-friendly license!”x.com
- SUPPORT@kwindla“handles overlapping speech well, can identify up to eight speakers, and pairs with any transcription model”x.com
- SOURCEmarketbrief.now
- SOURCEthe-decoder.com
- SOURCEmarktechpost.com