spurtcarl/finesub-diarizen-cpu

DiariZen (wavlm-large-s80-md-v2) as TorchScript — CPU build

Non-commercial use only (CC BY-NC 4.0). This is a format conversion made for FineSub's speaker diarization, so it can run without the diarizen and pyannote.audio packages. It is not a new model and was not retrained.

File	What it is	From	Licence
diarizen_segmentation.ts	WavLM-Large (pruned) + Conformer local segmentation, traced forward pass	BUT-FIT/diarizen-wavlm-large-s80-md-v2 @ 027b3e22	CC BY-NC 4.0
diarizen_embedding.ts	WeSpeaker ResNet34 speaker-embedding network (the part after fbank), traced	pyannote/wespeaker-voxceleb-resnet34-LM @ 837717dd	CC BY 4.0
plda/plda.npz, plda/xvec_transform.npz	PLDA and x-vector transform, copied unchanged	BUT-FIT/diarizen-wavlm-large-s80-md-v2	CC BY-NC 4.0

The bundle carries the most restrictive of the three: CC BY-NC 4.0.

CPU only

TorchScript records the device it was traced on. These files were traced on CPU and run only on CPU; the CUDA build is spurtcarl/finesub-diarizen-cuda.

What changed from the originals

• Traced with torch.jit.trace (torch 2.11.0+cu128); not frozen, not quantised, weights unchanged.
• The traced modules were checked against the eager originals bit for bit, repeated calls at batch sizes 1–3, before writing.
• Everything else DiariZen does — windowing, powerset decoding, VBx clustering — is reimplemented in FineSub; see its documentation for how that port was characterised.

Export script: tools/diarizen_export/export_diarizen.py in the FineSub repository.

Credit

DiariZen is by BUT Speech@FIT. Please cite their work (see the original model card) and the WeSpeaker paper (Wang et al., ICASSP 2023) for the embedding model. Some of DiariZen's training data is research-only, which is why its weights — and therefore this conversion — may not be used commercially.