johnpattern/safenotes-diar-splus-plda

pyannote community-1 PLDA (VBx transforms) for Safenotes

The PLDA / x-vector transforms of pyannote's speaker-diarization-community-1 pipeline, precomputed into the single binary file Safenotes' native VBx speaker clustering reads (no Python at run time).

This is a format conversion made by Safenotes (a local, on-device meeting notetaker) so the model runs on Intel hardware (Safenotes' native VBx clustering). 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

• pyannote speaker-diarization-community-1 by pyannote (pyannoteAI; Hervé Bredin and the pyannote.audio authors): pyannote/speaker-diarization-community-1 at revision 3533c8cf8e369892e6b79ff1bf80f7b0286a54ee, licensed CC-BY-4.0 (model card metadata). Its plda/ files (plda.npz, xvec_transform.npz) are, per the upstream plda/README.md, a "PLDA model trained by BUT Speech@FIT group", with thanks to Jiangyu Han and Petr Pálka for the integration of VBx in pyannote.audio. VBx: F. Landini, J. Profant, M. Diez, L. Burget, "Bayesian HMM clustering of x-vector sequences (VBx) in speaker diarization" (Computer Speech & Language, 2022).

This repository is distributed under the same licence: Creative Commons Attribution 4.0 International (CC-BY-4.0). The full licence text is in LICENSE.

Changes made by Safenotes

vbx-plda/plda_vbx.bin was computed from upstream plda/plda.npz and plda/xvec_transform.npz by Safenotes' engine/tools/convert-plda.py: the generalized eigenproblem of the PLDA is solved once (scipy eigh, exactly as pyannote.audio 4.0.7's VBx setup does) and the arrays mean1 (256), lda (256x128), mean2 (128), mu (128), tr (128x128) and psi (128) are written as little-endian float64 in Safenotes' SNPLDA1 format. Numerically the same transforms; no retraining. The upstream segmentation and embedding models are not included here.

How Safenotes uses it

Safenotes uses it for speaker separation after the meeting (VBx clustering of WeSpeaker embeddings), so users never need a Hugging Face login for the gated upstream repo. 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

1 files, 398,448 bytes.

file	bytes	sha256
vbx-plda/plda_vbx.bin	398,448	d1a496633687a2f008345cee013918e8258da06c28f4cf5bbc3df67fdd0ec626

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.