| --- |
| license: cc-by-4.0 |
| library_name: mlx |
| tags: |
| - mlx |
| - speaker-diarization |
| - witness |
| --- |
| |
| # witness-diarize-plda-redimnet2 |
|
|
| Big-corpus (921-speaker, LibriSpeech-refit) PLDA fit for the ReDimNet2-B6 |
| x-vector space, in the raw little-endian `*.f32` blob layout the witness VBx |
| backend loads (community-1 layout). Pairs with |
| `littlebearlabs/witness-redimnet2-b6-mlx` for VBx clustering. |
|
|
| ## Attribution |
|
|
| - **Fit**: ours — `.research/diarization/fit_resnet293_plda.py` + |
| `gen_plda_vbx_fixtures.py`, refit on a 921-speaker corpus in the |
| ReDimNet2-B6 192-d x-vector space. |
| - **Method / layout**: BUT VBx (variational Bayes HMM x-vector clustering) + |
| the pyannote `community-1` PLDA file layout. The ReDimNet2-B6 embedder these |
| x-vectors come from is MIT (PalabraAI); the VBx method and the community-1 |
| layout are the upstream references. |
| - **License**: CC-BY-4.0 (our fit; attribute witness + the BUT VBx / community-1 |
| lineage). |
|
|
| ## What's in this repo |
|
|
| Six raw little-endian f32 blobs (the full set the VBx backend loads): |
|
|
| - `transform_mean1.f32` (xvec_dim=192) |
| - `transform_lda.f32` (192·128) |
| - `transform_mean2.f32` (128) |
| - `plda_mu.f32` (128) |
| - `plda_tr.f32` (128·128) |
| - `plda_psi.f32` (128, the across-class covariance diagonal) |
|
|
| --- |
|
|
| *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`.* |
|
|
