NVIDIA open-sources Nemotron 3 Diarization, #1 on VoiceArena with 14.72% DER
solyarisoftware · x · 2026-09-24
NVIDIA released Nemotron 3 Diarization, an open-weight 100M-parameter speaker diarization model ranked #1 on VoiceArena's Diarization leaderboard with a 14.72% Diarization Error Rate (DER).
- Supports up to 8 speakers across 21 languages
- Works offline and online, with variable latency down to 0.32s
- 15000 RTFx throughput
- Licensed for both commercial and non-commercial use
The model identifies who spoke when—including overlapping speech—and pairs with ASR to produce speaker-attributed transcripts for meetings, call analytics, podcasts, and voice-agent memory.
Related event: NVIDIA Open-Sources Nemotron 3, Topping Speaker Diarization Benchmarks(10 posts)→
More from Models
- Leak Chatter: Frontier Model 8 Months Ahead of Astra, Only 2 Months Ahead of Trend — scaling01 · 2026-09-24
- Apple Open-Sources LensVLM-9B, a Qwen3.5-9B Finetune That Shrinks Long Docs into Page Images — victormustar · 2026-09-24
- OpenAI power users burn billions of tokens daily; Navier-Stokes project hits trillions of input tokens — scaling01 · 2026-09-24
- Navier-Stokes proof burned 130B output and trillions of input tokens, dwarfing top OpenAI users' few billion daily — scaling01 · 2026-09-24
- Meta Muse scores 4.88 across 40k ratings — analyst says product, distribution, price make it near-unbeatable — RihardJarc · 2026-09-24
- Audi 3D Modeling Benchmark Rerun: Only GPT-6 Astra and Opus 5.5 Are Worth Taking Seriously — kevinkern · 2026-09-24