gpillon/Nemotron-3-Diarization-GGUF-ignis

Nemotron 3 Diarization GGUF for ignis

Run it with ignis

This file is the diarizer of ignis, a local, OpenAI-compatible inference server: it tells the speakers of a conversation apart while ignis's speech model transcribes it, so a live transcript reads "Speaker 1, Speaker 2…". ignis runs it on the CPU, beside the speech model, with our own Rust port of the model: it takes no VRAM, and your audio never leaves the machine.

ignis-server model download nemotron-3-diarization

A Realtime transcription session asks for speakers with session.ignis.diarization: "live" (or "file" for audio fed faster than real time); each utterance's speaker arrives in ignis.input_audio_transcription.speaker before its completed, and a change of speaker ends an utterance. The diarizer loads at the first session that asks and unloads with the speech model.

Speakers come to ignis after v0.8.0 (tracking issue #361). Documentation: gpillon.github.io/ignis.

Files

File	What
Nemotron-3-Diarization-F16.gguf	the weights, F16 (norms, biases and the silence embedding F32), with the checkpoint's streaming configuration in the metadata

SHA-256: SHA256SUMS. The ignis catalog pins this repo's revision and this hash.

Measured with ignis

ignis's port against the checkpoint's own Transformers code (FP32, streamed chunk by chunk with the same streaming state): every 10 ms speaker probability within 6e-6.

Diarization error rate at a 0.25 s collar, twenty synthetic Italian conversations of 2 to 6 speakers built from VoxPopuli, each runtime's probabilities segmented by NeMo-Speech.cpp's post-processing:

geometry	ignis, this file	NeMo-Speech.cpp (Q8_0)
live: chunk 2 s, look-ahead 0.32 s (2.32 s of latency)	3.57 %	3.60 %
file: the card's offline geometry (30.4 s)	3.77 %	3.56 %

F16 scores as F32 does (3.57 %); Q8_0 in ignis's kernels scored 4.15 %.

Speed at the live geometry, a 297 s conversation:

	i9-10900K, 9 threads	4-core container, 3 threads
times real time	6.4x (0.32 s per 2 s step)	3.2x
share of the machine at real time		23 % of the four cores
beside Qwen3-ASR 1.7B transcribing the same session live	speakers within 3.6 s of the audio

On three AMI meetings (English, the headset mix) the error was 30 %, half the speech of one of them missed; the speakers it did hear were rarely confused (2 %).

How it was made

Converted by us from the official nvidia/Nemotron-3-Diarization at revision f667ed73aee57d40cc39428eb768b4fd87a0a29e (model.safetensors, sha256 c074d86335b3b794f8fa5edc25594558f128bdb3914d27806a3a5a2e44963cb6) with ignis's own converter, tools/speech-diarizer/convert.py --type F16 in the ignis repository: every matmul weight to F16, the upsampling convolution's taps reordered, the checkpoint's streaming_config written into the metadata. No weight was retrained or changed otherwise.

License and attribution

Nemotron 3 Diarization is by NVIDIA (nvidia/Nemotron-3-Diarization), released under the OpenMDW License Agreement, version 1.1 (LICENSE, a copy of the agreement, as it requires of a distribution). This file is a converted form of it, made by gpillon for ignis.