Nemotron 3 Diarization
An open-weight streaming speaker diarization model introduced by NVIDIA, based on the Streaming Sortformer architecture. It supports both streaming and offline inference, handles up to eight speakers, and is applicable to real-time meeting transcription and customer service audio analysis.
Timeline (last 90 days)
Nemotron 3 Diarization can be used together with NVIDIA's Parakeet series models, with Parakeet handling transcription and Nemotron handling speaker identification.
Nemotron 3 Diarization can run entirely locally, so audio does not need to be uploaded to the cloud.
Nemotron 3 Diarization is capable of distinguishing up to 8 different speakers in real-time streaming audio.
Nemotron 3 Diarization uses streaming frame-level inference, enabling real-time label output without waiting for an entire recording to finish.
Nemotron 3 Diarization has approximately 100 million parameters, making it a lightweight model.
Nemotron 3 Diarization has not yet officially disclosed specific engineering details such as recognition accuracy, latency performance, or hardware requirements.
NVIDIA released the Nemotron 3 Diarization speaker diarization model.
The Nemotron 3 Diarization model currently supports simultaneous tracking of 8 speakers.
The Nemotron 3 Diarization model is language agnostic, distinguishing speakers based on acoustic features rather than linguistic content.
The Diarization model and the ASR model are two separate models, rather than being packaged as a single unit.
2 more timeline events
All Facts (12)
Nemotron 3 Diarization can be used together with NVIDIA's Parakeet series models, with Parakeet handling transcription and Nemotron handling speaker identification.
50%UnverifiedNemotron 3 Diarization has approximately 100 million parameters, making it a lightweight model.
50%UnverifiedNemotron 3 Diarization is capable of distinguishing up to 8 different speakers in real-time streaming audio.
50%UnverifiedNemotron 3 Diarization uses streaming frame-level inference, enabling real-time label output without waiting for an entire recording to finish.
50%UnverifiedNemotron 3 Diarization can run entirely locally, so audio does not need to be uploaded to the cloud.
50%UnverifiedNemotron 3 Diarization has not yet officially disclosed specific engineering details such as recognition accuracy, latency performance, or hardware requirements.
50%UnverifiedThe Nemotron 3 Diarization model currently supports simultaneous tracking of 8 speakers.
50%UnverifiedThe Nemotron 3 Diarization model is language agnostic, distinguishing speakers based on acoustic features rather than linguistic content.
50%UnverifiedThe Diarization model and the ASR model are two separate models, rather than being packaged as a single unit.
50%UnverifiedNemotron 3 Diarization is an open-weight streaming speaker diarization model that supports both streaming and offline inference, can handle up to eight speakers, and outputs them in the order in which they first appear in the audio.
50%UnverifiedNemotron 3 Diarization employs the arrival-order speaker cache (AOSC) and FIFO queue mechanism introduced for Streaming Sortformer, with a single checkpoint supporting multiple latency configurations ranging from 80 ms input buffering to 30.4-second offline-style buffering, and output frame resolution adjustable in multiples of 10 ms.
50%UnverifiedNVIDIA released the Nemotron 3 Diarization speaker diarization model.
60%