Hugging Face's logo

FluidInference
/
speaker-diarization-coreml

Voice Activity Detection
Core ML
speech
audio
voice
speaker-diarization
speaker-change-detection
speaker-segmentation
Model card Files
xet
 Community
speaker-diarization-coreml / PROVENANCE.md
hanwenggggggg's picture
hanwenggggggg
Document Community-1 provenance and license scope
df2625a
preview code
|
Raw History Blame Contribute Delete
3.38 kB

Community-1 artifact provenance

This document covers the supported Community-1 artifacts in repository snapshot 1ed7a662fdc7109e36d822db793ee6eebdaf8594. Exact artifact hashes are recorded in provenance.json.

Source mapping

Converted artifact Upstream input
Segmentation segmentation/pytorch_model.bin
Embedding embedding/pytorch_model.bin (WeSpeaker ResNet34)
FBank Feature-extraction graph configured from the embedding checkpoint; no separate learned checkpoint
PLDA, PldaRho plda/plda.npz, plda/xvec_transform.npz
PLDA JSON resources The same two NPZ files, serialized as Float32 tensors

The immutable upstream reference snapshot is pyannote/speaker-diarization-community-1@3533c8cf8e369892e6b79ff1bf80f7b0286a54ee. Its relevant file history identifies checkpoint commit 5428423440de6baac451412694bb79a0ad644bea and PLDA import commit 7aa6699f402495053a3898e112874cb3e68ee064.

The PLDA files are byte-identical to the BUT Speech@FIT files with these hashes:

File SHA-256
plda.npz 9b77bcd840692710dd3496f62ecfeed8d8e5f002fd991b785079b244eab7d255
xvec_transform.npz 325f1ce8e48f7e55e9c8aa47e05d2766b7c48c4b25b8de8dd751e7a4cc5fbe8f

All six tensors in the two published JSON resources exactly match these inputs after conversion to Float32. The six constants derived by the published PLDA transformation also occur byte-for-byte in PldaRho.mlmodelc at the locations referenced by its MIL graph.

Recorded conversion environment

The modern artifacts record:

  • PyTorch 2.8.0
  • coremltools 9.0b1
  • TorchScript source dialect
  • coremlc 3500.32.1
  • MIL component 3500.14.1

The public conversion reference is Mobius commit 33fd6eab634966ae7db4d73da3376a90379642fb, including conversion and compilation scripts plus the locked Python environment. That commit postdates the artifact conversion dates, so it is a reference implementation rather than an attestation of the exact historical build.

Reproducible successor pipeline

Mobius commit ffbc3c8cae2d0ac83912a005a25a2874ced98a3a adds a release command that:

  1. Rejects mutable or abbreviated upstream revisions.
  2. Downloads material inputs from the selected immutable revision.
  3. Embeds source paths, hashes, upstream revision, and converter revision in the generated Core ML packages and JSON resources.
  4. Compiles the packages and emits provenance.json containing every input and output SHA-256 plus the build environment and commands.
  5. Refuses release builds from a dirty converter checkout.

Example:

uv run python build-release.py \
  --upstream-revision 3533c8cf8e369892e6b79ff1bf80f7b0286a54ee \
  --work-dir ./build/community-1-3533c8cf \
  --release-dir ./build/release-3533c8cf \
  --selective-fp16

Access to the gated upstream model and acceptance of its user conditions are required.

Historical limitations

  • The exact upstream checkout used for the original 2025 segmentation and embedding exports was not recorded.
  • The exact historical converter commit, invocation, Xcode installation, and SDK were not recorded.
  • The existing segmentation and embedding artifacts therefore remain historically reconstructed, not build-attested.
  • The older pyannote_segmentation and wespeaker* artifacts are outside this document's scope.