spurtcarl/finesub-speaker-embedding

WeSpeaker ResNet34 as TorchScript, with its PLDA

A format conversion made for FineSub's speaker diarization, so it can run without pyannote.audio. It is not a new model and was not retrained.

File	What it is	From	Licence
wespeaker-cuda.ts	WeSpeaker ResNet34 speaker-embedding network (the part after fbank), traced on CUDA	pyannote/wespeaker-voxceleb-resnet34-LM @ 837717dd	CC BY 4.0
wespeaker-cpu.ts	The same network, traced on CPU	same	CC BY 4.0
plda/plda.npz, plda/xvec_transform.npz	PLDA and x-vector transform, byte-identical copies	pyannote/speaker-diarization-community-1 @ 3533c8cf	CC BY 4.0

Two builds

TorchScript records the device it was traced on, so each file runs only on its own device. Both are here because together they are about 54 MB.

What changed from the originals

• Traced with torch.jit.trace (torch 2.11.0); not frozen, not quantised, weights unchanged.
• The traced modules were checked against the eager original bit for bit, repeated calls at batch sizes 1–3, before writing.
• The PLDA files are copied unchanged.

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

Credit

WeSpeaker: Wang et al., "WeSpeaker: A research and production oriented speaker embedding learning toolkit", ICASSP 2023; the checkpoint as packaged by pyannote. The PLDA is pyannote's, from speaker-diarization-community-1 (Bredin, pyannote.audio). Please credit them, not this repository.