Recogment/nemotron-3-diarization-onnx

Nemotron 3 Diarization — streaming step (ONNX) — Recogment mirror

Streaming speaker diarization for up to 8 speakers. This repo packages an ONNX export of one streaming step of the upstream NVIDIA checkpoint, so runtimes that don't depend on PyTorch / NeMo can use it.

⚠️ Unofficial. This is a community-maintained mirror of a derived artefact. For the canonical model, its evaluation and training details see nvidia/nemotron-3-diarization.

Files

File	Purpose	Size
nemotron-3-diarization.onnx	One stateless streaming step, fp32	~398 MB

The step graph

	Name	Shape
input	chunk_features	[1, T, 128] — 10 ms log-mel frames of the chunk plus its look-ahead; T a multiple of 8
input	cached_embeds	[1, C, 512] — arrival-order speaker cache ++ FIFO (C may be 0)
output	logits	[1, (C + T/8) · 8, 8] — 10 ms speaker logits for every input frame
output	chunk_embeds	[1, T/8, 512] — the chunk's encoder-input embeddings, which join the FIFO

The speaker cache, the FIFO and the cache compression are not in the graph: the caller runs them (a Rust port of transformers' Nemotron3DiarizationSpeakerCache lives in parakeet-rs as nemotron_diar.rs). ONNX metadata carries the streaming geometry (chunk 340 / look-ahead 40 / FIFO 40 / cache update period 300 encoder frames), the cache policy constants, and the model's learned silence embedding (silence_embeds, 512 × f32 little-endian, hex).

Provenance

Exported from the upstream Hugging Face checkpoint of nvidia/nemotron-3-diarization through its transformers implementation (scripts/export_nemotron_diar.py in parakeet-rs, opset 17). No quantization, retraining, or fine-tuning. Checked against the PyTorch reference over whole recordings: identical per-step logits to ~1e-4 until float noise flips a near-tied cache-compression choice; a Rust driver of this graph matches PyTorch streaming to a median |Δp| of 3.5e-9 with no p > 0.5 decision differing over a 30-minute call.

Why this mirror exists

The Recogment daemon ships with a pinned set of model SHA-256 hashes and a zero-egress sandbox; the only outbound network in the whole product is a companion downloader that fetches from public HTTPS sources. This mirror lets that downloader fetch the diarizer when a user selects it.

If you're not building Recogment, you almost certainly want the official NVIDIA repo linked above instead.

License

This work is licensed under the OpenMDW License Agreement 1.1 — see LICENSE.txt in this repo, and NOTICE.txt for the origin notice. Attribution: NVIDIA Corporation, via nvidia/nemotron-3-diarization.