| --- |
| license: cc-by-4.0 |
| base_model: |
| - nvidia/diar_streaming_sortformer_4spk-v2 |
| library_name: onnx |
| tags: |
| - onnx |
| - speaker-diarization |
| - sortformer |
| - streaming |
| - safenotes |
| --- |
| # Streaming Sortformer 4spk v2 (ONNX) for Safenotes |
|
|
| NVIDIA's Streaming Sortformer speaker-diarization model (up to 4 speakers) exported to ONNX with its streaming inputs (chunk, speaker cache, FIFO), plus the mel filterbank and window tables of its preprocessor. |
|
|
| This is a **format conversion made by Safenotes** (a local, on-device meeting notetaker) so the model runs on Intel |
| hardware (ONNX Runtime / sherpa-onnx). It is not an official release of |
| the original authors, who do not endorse it. All credit for the model goes to them. |
|
|
| ## Original model and attribution |
|
|
| - **Streaming Sortformer 4spk v2** by NVIDIA: [nvidia/diar_streaming_sortformer_4spk-v2](https://huggingface.co/nvidia/diar_streaming_sortformer_4spk-v2) at revision [`84edd514b8ef68004c10086918cd62f2148cbd59`](https://huggingface.co/nvidia/diar_streaming_sortformer_4spk-v2/tree/84edd514b8ef68004c10086918cd62f2148cbd59), licensed [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/legalcode) (model card metadata `license: cc-by-4.0` and its Licence section at that revision). Note: the later v2.1 model is under the NVIDIA Open Model License and is NOT used here. |
|
|
| This repository is distributed under the same licence: [Creative Commons Attribution 4.0 International (CC-BY-4.0)](https://creativecommons.org/licenses/by/4.0/legalcode). The full licence text is in [`LICENSE`](LICENSE). |
|
|
| ## Changes made by Safenotes |
|
|
| `own/sortformer-v2.onnx`: the original .nemo checkpoint (sha256 b371afce2c49…) exported by Safenotes with torch.onnx (opset 18, fp32), with the streaming state as explicit inputs/outputs (the state update runs outside the graph, as in NeMo's forward_streaming). `mel_fb.npy` / `mel_window.npy`: the mel filterbank and window of the model's own preprocessor, saved as arrays. No retraining. Parity with NeMo 3.0.0 forward_streaming: 0 frames flipped. |
|
|
| ## How Safenotes uses it |
|
|
| Safenotes uses it for speaker labels during the call (ONNX Runtime on the CPU). The Safenotes app downloads the files below from a pinned revision of this repository and checks every |
| file's sha256 before using it. Everything runs on the user's PC; no audio or text leaves it. |
|
|
| ## Files |
|
|
| 3 files, 492,379,127 bytes. |
|
|
| | file | bytes | sha256 | |
| |---|---:|---| |
| | `mel_fb.npy` | 131,712 | `462cb0ddf0ee659b44e73eb67f06d1ebb4663a16d5a2fa98315ad875955424c1` | |
| | `mel_window.npy` | 1,728 | `39daf4ff0914631de0068adb484e81f390e8c053c836ad5ba92f76849bf334fd` | |
| | `own/sortformer-v2.onnx` | 492,245,687 | `a5974a8d8689bf5d7f03167bf290cd5a1a81081cb2b14a26ecae3fac418812f4` | |
|
|
| ## No warranty |
|
|
| Provided "as is", without warranties or conditions of any kind, express or implied, including fitness for a particular |
| purpose and accuracy of the output (see the licence). Use at your own risk. |
|
|
