Streaming Transformer-based speaker diarization, fixed at 4 simultaneous speakers per chunk. This repo packages ONNX exports of the upstream NVIDIA NeMo checkpoint so that runtimes which don't depend on PyTorch / NeMo can use the model directly.
Both files contain the same model weights — the difference is only in the ONNX graph's batch / time-axis declarations. The dynamic variant is what you use for sub-chunk streaming; the static variant gives onnxruntime more room to optimise when you have a fixed 30s window.
These ONNX files were exported from the upstream NeMo
.nemo checkpoint of
nvidia/diar_streaming_sortformer_4spk-v2. No quantization, retraining, or fine-tuning — weights round-trip bit-for-bit through the export.
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 binary that fetches from public HTTPS sources. Until this mirror existed, the Sortformer files had to be shipped to beta testers out-of-band as a tarball. With this mirror in place, the downloader can fetch them automatically alongside the rest of the public catalogue.
If you're not building Recogment, you almost certainly want the official NVIDIA repo linked above instead — it includes the original .nemo checkpoint, configuration, and reference NeMo inference code.
This work is licensed under
CC-BY-4.0, matching the upstream NVIDIA repo. Attribution: NVIDIA Corporation, via
nvidia/diar_streaming_sortformer_4spk-v2.
1@misc{nvidia2025sortformerv2,
2 title={Streaming Sortformer Diarizer 4spk v2},
3 author={NVIDIA},
4 year={2025},
5 publisher={Hugging Face},
6 howpublished={\url{https://huggingface.co/nvidia/diar_streaming_sortformer_4spk-v2}}
7}
