Nemotron 3
An open-weight speaker diarization model released by NVIDIA with approximately 100 million parameters. It supports recognition of up to eight speakers and operates in both streaming and offline modes. The model is released under a license that permits commercial use.
Timeline (last 90 days)
NVIDIA's Nemotron speaker diarization model has only 99M parameters, supports both streaming and offline inference, handles up to 8 speakers, and achieves a minimum streaming latency of 80 milliseconds.
The Nemotron 3 speaker diarization model has approximately 100 million parameters and requires only 4 GB of VRAM to run.
Nemotron 3 can handle up to 8 speakers, whereas the upper limit for most past models was typically only 3 to 4.
The Nemotron 3 speaker diarization model is designed to work in a decoupled manner with any speech recognition model and can be used alongside ASR models such as Parakeet.
Nemotron 3 can be seen as a direct replacement for SortFormer and performs significantly better.
The NVIDIA Nemotron 3 speaker diarization model is an open-weight model and adopts a license that permits commercial use.
In the author's local deployment tests, transcription and diarization of an audio clip of one and a half minutes took only three to four seconds.
The model completed all processing for a full episode of the All-In podcast lasting over one hour in approximately 148 seconds, and output word-for-word transcriptions and statistics for each speaker.
The Nemotron 3 speaker diarization model itself only outputs labels such as speaker_1 and speaker_2 and does not automatically identify names.
Nemotron 3 supports both streaming and offline modes and is language-agnostic.
All Facts (8)
NVIDIA's Nemotron speaker diarization model has only 99M parameters, supports both streaming and offline inference, handles up to 8 speakers, and achieves a minimum streaming latency of 80 milliseconds.
50%UnverifiedThe Nemotron 3 speaker diarization model has approximately 100 million parameters and requires only 4 GB of VRAM to run.
50%UnverifiedNemotron 3 can handle up to 8 speakers, whereas the upper limit for most past models was typically only 3 to 4.
50%UnverifiedThe Nemotron 3 speaker diarization model is designed to work in a decoupled manner with any speech recognition model and can be used alongside ASR models such as Parakeet.
50%UnverifiedThe NVIDIA Nemotron 3 speaker diarization model is an open-weight model and adopts a license that permits commercial use.
50%UnverifiedNemotron 3 supports both streaming and offline modes and is language-agnostic.
50%UnverifiedThe Nemotron 3 speaker diarization model itself only outputs labels such as speaker_1 and speaker_2 and does not automatically identify names.
50%UnverifiedNemotron 3 can be seen as a direct replacement for SortFormer and performs significantly better.
50%