NVIDIA open-sources Nemotron 3 Diarization, tops diarization benchmark by 24%
On September 23, NVIDIA released and open-sourced Nemotron 3 Diarization, a speaker diarization model with just 100M parameters that can track "who spoke when" in multi-speaker overlapping speech. It supports streaming output, tracks up to 8 speakers, is available on Hugging Face, and ships with a commercially friendly license. The model ranked first among 12 systems in the inaugural Diarization-Bench on voicearena.ai, with a diarization error rate of 14.72%—about 24% lower than the runner-up.
Confirmed
- Nemotron 3 Diarization was officially released and open-sourced by NVIDIA, is available on Hugging Face, has 100M parameters, and supports up to 8 speakers.
- Official benchmark results: ranked first among 12 systems in the first Diarization-Bench on voicearena.ai, with an error rate of 14.72%, about 24% lower than second place.
- The model outputs speaker labels (not transcriptions) and supports streaming.
Developer tests and reactions
- @futterneid tested it locally on a voice agent using 1-second streaming chunks, saying it solves the "who is speaking" pain point and can be used to let a bot remember different speakers (new people).
- @andimarafioti noted that a key gap in today's voice agents is precisely the inability to distinguish speaker identity, which various phone-call demos obscure; Nemotron 3 Diarization's commercially friendly license makes it well suited for real-world deployment.
Why it matters
- Speaker diarization has long been a weak spot for real-time voice agents (meeting transcription, customer service, home robots, etc.), and prior open-source solutions have performed poorly on overlapping speech.
- The combination of a small parameter count (100M), streaming output, and a commercially friendly license means low barriers to local deployment—individual developers (as @futterneid's practice shows) can plug it directly into their own voice agent pipelines.
2026-09-23 ~ 2026-09-23 · 6 related posts
- Episode 1: NVIDIA open-sources Nemotron 3 Diarization, tops diarization benchmark by 24%(2026-09-23, 6 posts)
- Episode 2: NVIDIA Nemotron 3 Diarization Launches on Baseten with Day-0 Support(2026-09-24, 2 posts)
Primary sources
- NVIDIA releases Nemotron 3 Diarization model handling up to 8 overlapping speakers with 100M params — NVIDIAAI ·
- Nemotron 3 Diarization tops Diarization-Bench with 14.72% error rate, ~24% lower than runner-up — NVIDIAAI ·
- Nemotron 3 Diarization in action: streaming 8-speaker tracking powers a memory-equipped robot — futterneid ·
- [source] NVIDIA releases Nemotron 3 Diarization model handling up to 8 overlapping speakers with 100M params — NVIDIAAI · 2026-09-23
- [source] Nemotron 3 Diarization tops Diarization-Bench with 14.72% error rate, ~24% lower than runner-up — NVIDIAAI · 2026-09-23
- NVIDIA open-sources Nemotron 3 Diarization to fix voice agents' speaker-blindness — andimarafioti · 2026-09-23
- [source] Nemotron 3 Diarization in action: streaming 8-speaker tracking powers a memory-equipped robot — futterneid · 2026-09-23
- NVIDIA Releases Nemotron 3 Diarization, Massively Outperforming Rivals — Sam Witteveen · 2026-09-23
1 near-duplicate retellings: NVIDIAAI