Not established from the available sources.
NVIDIA · MODEL RELEASE TRACKER
Nemotron 3 Diarization
Nemotron 3 Diarization is a Nvidia-released diarization model (~100M parameters) that identifies which speaker is talking in recordings or live audio, supports up to eight speakers, detects overlapping speech, and whose weights are freely available.CURRENT SNAPSHOT4/5 DIMENSIONS WITH DATA
The dimensions that change the decision.
Audio buffer settingsAudio buffer configurable at four levels ranging from 30.4 down to 0.32 seconds; shorter buffers generally reduce accuracy.
Primary taskSpeaker diarization (who spoke when) — audio diarization
Input modalitiesWorks with both recordings and live audio (streaming).
Benchmark performanceRanks #1 on Voice Arena's Diarization-Bench with a 14.72% Diarization Error Rate (DER).
Benchmark performanceRanked first on the Diarization-Bench (VoiceArena) with a 14.72% error rate; the next best system had 19.3%. The benchmark counts overlapping speech and small misalignments as errors.
Weights availabilityModel weights are freely available.
RELEASE TIMELINE
Published, source-backed release events only.
Nvidia releases Nemotron 3 Diarization, a 100M-parameter real-time speaker diarization model
Nvidia released Nemotron 3 Diarization, a ~100 million-parameter model whose weights are freely available. The model can identify up to eight speakers (including overlapping speech), works on live and recorded audio with configurable buffer sizes, and achieves a 14.72% error rate on VoiceArena's Diarization-Bench, outperforming the prior best system and reducing error versus Streaming Sortformer by about 41% on certain tests.
NVIDIA releases Nemotron 3 Diarization — open-weight 100M model for real-time multi-speaker diarization
NVIDIA published Nemotron 3 Diarization, an open-weight, 100M-parameter speaker-diarization model that ranks #1 on Voice Arena's Diarization-Bench (14.72% DER). The model supports up to eight anonymous speaker channels, handles overlapping speech in streaming and offline modes, and uses arrival-order speaker caching (AOSC) and a FIFO context buffer; training included public and licensed data, with David AI data reducing compound DER by 0.77 points.
VERIFIABLE FACTS
Every value stays attached to a source and date.
MODALITIES · Primary taskDEVELOPER CLAIM
Speaker diarization (who spoke when) — audio diarization
CAPABILITIES · Parameter countDEVELOPER CLAIM
About 100 million parameters.
CAPABILITIES · Max supported speakersDEVELOPER CLAIM
Supports up to eight speakers in live and recorded conversations
CAPABILITIES · Notable capabilitiesDEVELOPER CLAIM
Handles overlapping speech; supports chunked processing for flexible recording lengths; customizable streaming latency; orders output speakers by first appearance (Sortformer approach).
AVAILABILITY · Weights availabilityDEVELOPER CLAIM
Model weights are freely available.
BENCHMARKS · Benchmark performanceDEVELOPER CLAIM
Ranks #1 on Voice Arena's Diarization-Bench with a 14.72% Diarization Error Rate (DER).
RELEASE · Release announcementDEVELOPER CLAIM
Reportedly released by Nvidia (article published 2026-09-27).
CAPABILITIES · Maximum number of speakersDEVELOPER CLAIM
Can tell apart up to eight speakers.
CAPABILITIES · Overlapping speech detectionDEVELOPER CLAIM
Can detect when multiple people talk at the same time (overlapping speech).
MODALITIES · Input modalitiesDEVELOPER CLAIM
Works with both recordings and live audio (streaming).
CONTEXT WINDOW · Audio buffer settingsDEVELOPER CLAIM
Audio buffer configurable at four levels ranging from 30.4 down to 0.32 seconds; shorter buffers generally reduce accuracy.
BENCHMARKS · Benchmark performanceDEVELOPER CLAIM
Ranked first on the Diarization-Bench (VoiceArena) with a 14.72% error rate; the next best system had 19.3%. The benchmark counts overlapping speech and small misalignments as errors.
WHAT CHANGED
Stored passport versions, without reconstructed history.
CONTEXT WINDOW · Audio buffer settings+ Audio buffer configurable at four levels ranging from 30.4 down to 0.32 seconds; shorter buffers generally reduce accuracy.
BENCHMARKS · Benchmark performance+ Ranked first on the Diarization-Bench (VoiceArena) with a 14.72% error rate; the next best system had 19.3%. The benchmark counts overlapping speech and small misalignments as errors.
MODALITIES · Input modalities+ Works with both recordings and live audio (streaming).
CAPABILITIES · Maximum number of speakers+ Can tell apart up to eight speakers.
CAPABILITIES · Overlapping speech detection+ Can detect when multiple people talk at the same time (overlapping speech).
CAPABILITIES · Parameter count100M parameters→About 100 million parameters.
RELEASE · Release announcement+ Reportedly released by Nvidia (article published 2026-09-27).
AVAILABILITY · Weights− Open-weight (weights released / open)
AVAILABILITY · Weights availability+ Model weights are freely available.