Nemotron-3-Diarization — ONNX

ONNX export of nvidia/Nemotron-3-Diarization (HF Transformers implementation), license OpenMDW-1.1 (see LICENSE; attribution to NVIDIA).

Streaming Sortformer, up to 8 speakers, 16 kHz mono input, 10 ms output frames, sigmoid activity per speaker.
Folders int8/ (dynamic QInt8 MatMul quantization, recommended, ~105 MB) and fp32/ (reference, ~399 MB); each is self-contained.
Three files per variant: diar_frontend.onnx (log-mel, no STFT op), diar_step*.onnx (stateless transformer step), diar_config.json (geometry, speaker-cache constants, silence embedding). The speaker-cache (AOSC) update is host code: see SpeechLib.Diarization in the VoiceType repo.
int8 DER is within ±0.5 absolute of fp32 on the conversion test sets.
Converted with tools/converters/NemotronDiarization/export_onnx.py (torch 2.14.1, transformers 5.19.0.dev0, onnx 1.23.2, opset 17).
Tested with ONNX Runtime 1.30 CPU (win-x64). Not verified on phones.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for DimQ1/nemotron-3-diarization-onnx

Base model

Quantized

(33)

this model