Nvidia publishes open weights for a speaker-labelling model that runs live

reported by 2 outlets· 2 articles · confidence: high · first seen 2026-09-23 00:00 UTC

What this means for you

If you pay a vendor per minute for speaker labels, this is now self-hostable: 100M parameters, eight speakers, one checkpoint for live and recorded audio. Check the licence before shipping, since the sources do not state it, and re-measure on your own audio — the leaderboard score carries no evaluation date.

Nvidia published Nemotron 3 Diarization on Hugging Face on 23 September: a 100M-parameter open-weight model that marks who spoke when in a conversation, a job usually run alongside speech recognition rather than inside it. It handles up to eight speakers, including overlapping speech, and a single checkpoint serves both recorded files and live audio with adjustable streaming latency. Nvidia reports a 14.72% diarization error rate and first place on VoiceArena's Diarization-Bench; no evaluation date is given. Its earlier streaming model, Sortformer, covered four speakers. The licence terms are not stated in the material available.

Key facts

  • ·Nvidia released Nemotron 3 Diarization, a 100M-parameter open-weight speaker diarization model, on Hugging Face on 23 September 2026. source
  • ·The model supports up to eight speakers, including overlapping speech, across live and recorded conversations. source
  • ·One checkpoint handles both offline recordings and real-time streaming. source
  • ·Nvidia reports a 14.72% Diarization Error Rate and first place on VoiceArena's Diarization-Bench leaderboard, with no evaluation date given. source
  • ·Nvidia's earlier Streaming Sortformer model covered four speakers; the diar_streaming_sortformer_4spk-v2.1 checkpoint is used as the baseline. source

What the sources say

  • Hugging Face Blog — Technical release post covering the architecture, streaming memory handling, and the leaderboard score against an earlier baseline.
  • MarkTechPost — Short news item confirming the release, the eight-speaker limit and one checkpoint for both modes.

Sources

The original reporting. Follow these — they did the work.

← the wire