Sona speaker model: Nemotron 3 Diarization, one streaming step, int8

The speaker model of Sona, an on-device Swedish transcription app for Android. It is an int8 build of the ONNX export of one streaming step of NVIDIA's Nemotron 3 Diarization made by beshkenadze (revision 1e1e5e2993f9e9558c10a8a72337c769f0ec0516), which in turn is a conversion of NVIDIA's checkpoint (revision f667ed73aee57d40cc39428eb768b4fd87a0a29e).

The file is reproducible bit for bit with research/diarization-benchmark/quantize.py and the pinned requirements.txt there. The graph's inputs and outputs are unchanged from the fp32 export: a 30.4 s chunk of log-mel frames plus the speaker cache (264 frames) and FIFO (40), out come per-speaker probabilities every 80 ms and the chunk's embeddings. The mel front end and the speaker-cache update (NeMo's streaming_update_async) run outside the graph; Sona's are in :core:diarization, checked against a Python reference in the same repository. The learned silence embedding the cache update needs (learnable_sil_emb.f32) is in beshkenadze's repo.

On a Swedish test set built from public data (Riksdagen debates labelled with RixVox-v2, and simulated conversations), this build matches fp32 within noise; see the benchmark README.

License and attribution

Nemotron 3 Diarization is by NVIDIA Corporation and is released under the OpenMDW License Agreement, version 1.1; its text is in LICENSE, and it applies to these files as a conversion of that model. The ONNX export of the streaming step is by beshkenadze, under the same agreement. See the original model card for the model's training data, evaluation, intended use and limitations, all of which apply here unchanged.