| license: mit | |
| library_name: mlx | |
| tags: | |
| - mlx | |
| - speaker-diarization | |
| - witness | |
| # witness-pyannote-seg-mlx | |
| `pyannote/segmentation-3.0` powerset speaker segmentation, converted to a | |
| gate-free MLX safetensors for witness's on-device speaker diarization. | |
| ## Upstream attribution | |
| - **Model**: pyannote `segmentation-3.0` (PyanNet powerset segmentation). | |
| - **Authors / source**: pyannote — `pyannote/segmentation-3.0`. | |
| - **License**: MIT (the segmentation-3.0 model weights are MIT-licensed). | |
| ## What's in this repo | |
| - `model.safetensors` — gate-free converted weights produced by | |
| `.research/diarization/gen_seg_fixture.py`. Validated to 100% per-frame | |
| powerset-argmax agreement with the upstream export (logits max_abs ≈ 2.5e-5). | |
| --- | |
| *Converted to MLX for [witness](https://github.com/littlebearlabs/witness), an | |
| open-source Rust toolkit for on-device system capture on macOS. Generated by | |
| `.research/diarization/publish_weights.sh`.* | |