Nemotron-3-Diarization-preview
nvidia/Nemotron-3-Diarization-preview
Official · BF16 · nemo
Runtime certified: NeMo
| Precision / quantization | Format | Bits | Publisher trust | Deployment state |
|---|---|---|---|---|
| BF16 | nemo | 16 | Official | Runtime certified |
NVIDIA Nemotron · Speaker diarization · Nemotron 3
Nemotron 3 Diarization Preview: Specs, GPU Requirements and License
Nemotron 3 Diarization Preview is NVIDIA’s 100M-parameter speaker-diarization model for identifying who spoke when in streaming or recorded audio.
Total parameters100M
Active parametersDense model
Context windowN/A / Not directly comparable
Release statusPreview · Recent
ArchitectureStreaming Sortformer Transformer
ModalitiesAudio
Licenseother
Deployment stateTechnically runnable for evaluation
Commercial stateUnknown / Not Recommendable
Release date2026-08-27
Technical profile
| Field | Value |
|---|---|
| Purpose | Speaker diarization: identify who spoke when |
| Speaker capacity | Up to 8 speakers, ordered by arrival |
| Inference modes | Streaming and offline |
| Input | 16 kHz mono WAV, FLAC, OPUS or MP3 |
| Output | [T, 8] speaker activity probabilities |
| Input buffer | 80 ms to 30.4 s; chunked inference has no fixed duration limit |
| Runtime | NVIDIA NeMo with PyTorch |
| Precision | BF16 |
| Reference hardware | NVIDIA Blackwell RTX PRO 5000 |
Architecture, strengths and use cases
This gated, self-hostable preview belongs to the Nemotron Audio/Speech domain and Speaker Diarization track. It accepts 16 kHz mono audio and returns activity probabilities for as many as eight speakers ordered by arrival.
The 100M-parameter model uses a 31-layer Streaming Sortformer Transformer with rotary position embeddings. Ten-millisecond Mel frames are downsampled to 80 ms encoder frames, then a Conv1D stage produces speaker activity probabilities.
Streaming and offline diarization
Up to eight arrival-ordered speakers
Official NeMo checkpoint with BF16 evidence
Evaluation of meeting and call speaker segmentation
Streaming speaker-change analysis
Offline diarization research on recorded audio
Certified checkpoints and quantizations
Only exact, source-certified variants attached to this release are shown.
Recommended hardware
Use the exact gated .nemo checkpoint through NVIDIA NeMo with the documented PyTorch backend and BF16 precision. The NVIDIA reference measurement used an RTX PRO 5000 Blackwell GPU; this is test hardware, not a commercial hosting offer.
Unknown / Not Recommendable
Commercial rights are not verified, so this release is not commercially recommendable.
License and self-hosting considerations
Access requires acceptance of NVIDIA’s evaluation license. Input must be 16 kHz mono audio; chunked inference supports long recordings, while the buffer can be configured from 80 ms to 30.4 seconds.
The NVIDIA Software and Model Evaluation License permits internal testing and evaluation only. It prohibits production use and commercial hosting, and restricts use to NVIDIA GPU systems. Consult the official terms for the complete obligations.
Unknown / Not Recommendable. Commercial rights are not verified, so this release is not commercially recommendable.
Release timeline
nvidia/nemotron/nemotron-3-diarization-preview
Source details
Source revision: a43022e40b9883fb75078c47c5d4c995436863bb26e26f13a1a7e836ea519229
Frequently asked questions
What does Nemotron 3 Diarization Preview do?
It identifies speaker activity over time and assigns activity probabilities to as many as eight speakers in arrival order.
Can it process live and recorded audio?
Yes. NVIDIA documents both streaming and offline inference through the NeMo checkpoint.
Can it be used for commercial production?
No commercial offer is shown. The verified NVIDIA license limits this release to internal testing and evaluation and prohibits production use.
Official references
Last verified: 2026-09-03