Ever read a meeting transcript where every line is correct but you can't tell who said what? NVIDIA just fixed that — and gave the fix away for free.
On September 23, NVIDIA released Nemotron 3 Diarization, an open-weight, 100-million-parameter model that answers one of audio AI's most annoying questions: "who spoke when," in live or recorded audio.
It handles up to 8 simultaneous speakers with speaker-activity probabilities at 10-millisecond resolution — granular enough to catch rapid-fire crosstalk and interruptions. Streaming latency profiles go as low as about 320 ms, and there's an offline mode for full recordings.
The model ranked #1 on VoiceArena's Diarization-Bench leaderboard with a 14.72% diarization error rate. It was trained on roughly 10,000 hours of real conversations plus 82,611 hours of simulated multi-talker mixtures.
Under the hood is a neat engineering choice: a single-pass end-to-end design NVIDIA calls the "Arrival-Order Speaker Cache" (AOSC). Traditional diarization pipelines need multiple stages — segmenting audio, extracting speaker embeddings, then clustering them. Nemotron 3 collapses all of that into one forward pass, which makes deployment much simpler.
It's shipped under NVIDIA's OpenMDW 1.1 license on Hugging Face (nvidia/Nemotron-3-Diarization), through NVIDIA's NeMo framework, and via inference providers like Baseten and DeepInfra — with deployment reportedly as cheap as about $0.01 per audio hour in some configurations.
One important note: the model does NOT transcribe words and does NOT identify real people. It assigns anonymous labels (speaker_1, speaker_2...) based on the order they first appear. Pair it with a speech-recognition model and you get full speaker-attributed transcripts.
Why it matters: speaker attribution is the missing layer behind good meeting notes, podcast editing, call analytics, and voice agents. Every transcript app now has an open, state-of-the-art upgrade path — and at 100M parameters, it's cheap enough to run at scale. This is infrastructure for the voice-agent boom, not a demo.
Our take: NVIDIA keeps doing the quiet, unglamorous work the whole industry then builds on. Expect to see this inside products within months.
For more AI news and tool launches: https://aipost.tech
Comments
Post a Comment