Provenance and licensing details for diarization Core ML artefacts at `1ed7a662`

[mcironbrian-gh]

Hello, and thank you for making FluidAudio available.

We are conducting commercial due diligence for possible redistribution in an offline application. For the speaker-diarization-coreml repository at revision 1ed7a662fdc7109e36d822db793ee6eebdaf8594, could you please clarify:

• The exact upstream checkpoint names and revisions used for the segmentation, FBank, and embedding models.
• The source and licence basis of PldaRho.mlmodelc and plda-parameters.json.
• The conversion script or commit used to produce the Core ML artefacts, including relevant conversion-tool versions.
• Whether the repository’s CC-BY-4.0 licence is intended to cover those exact converted artefacts, including PldaRho.mlmodelc and plda-parameters.json.

Public links to the relevant upstream sources, licence notices, and reproducible conversion materials would be especially helpful. Thank you.

[Alex-Wengg]

Thanks for the precise revision. Here is what can be established for FluidInference/speaker-diarization-coreml at 1ed7a662fdc7109e36d822db793ee6eebdaf8594, with the remaining provenance gaps called out explicitly.

1. Segmentation, FBank, and embedding sources

The published conversion source loads the Community-1 segmentation and embedding checkpoints from a local copy of pyannote/speaker-diarization-community-1:

Converted artifact	Input in the Community-1 repository
Segmentation.mlmodelc	segmentation/pytorch_model.bin
Embedding.mlmodelc	embedding/pytorch_model.bin, the WeSpeaker ResNet34 backend
FBank.mlmodelc	Feature-extraction code configured from the embedding checkpoint's hyperparameters; there is no separate learned FBank checkpoint

An immutable upstream reference snapshot is available. However, the conversion script accepts a local directory and does not record its Git revision. The reviewed release metadata does not establish the exact upstream checkout used for each export, so that reference revision should not be presented as a certified build input.

2. PLDA source and license basis

The PLDA conversion module loads plda/plda.npz and plda/xvec_transform.npz from Community-1 and constructs the whitening/projection and rho transform. The JSON resources identify those same source paths.

Those parameter files originate from BUT Speech@FIT. Its PLDA-specific README describes their training on VoxCeleb2 development data, and its PLDA-specific license expressly covers both files under CC-BY-4.0, including commercial use. This is a separate license from that repository's main DiariZen model.

A static comparison verified that all six tensors in the pinned FluidInference JSON resources exactly match the BUT files after conversion to float32. Reconstructing the PLDA constants with the published transformation also produced six parameter arrays present byte-for-byte in the pinned PldaRho.mlmodelc/weights/weight.bin.

The explicit BUT PLDA license notice was added on 2026-08-31, after the pinned FluidInference snapshot. It provides a current upstream statement for those matching files; it does not document the permissions recorded at the original conversion date.

3. Conversion materials and tool versions

Public conversion materials are available at Mobius commit 33fd6eab634966ae7db4d73da3376a90379642fb, including conversion and compilation scripts, PLDA code, pyproject.toml, and uv.lock.

The pinned Community-1 artifacts record PyTorch 2.8.0, coremltools 9.0b1, and the TorchScript source dialect in their metadata. Their MIL build information records coremlc 3500.32.1 and MIL component 3500.14.1; see, for example, PldaRho metadata and MIL. The source environment additionally pins Python 3.10.12 and pyannote-audio 4.0.0.

The public source commit postdates the artifact conversions. It is a conversion reference, not a verified record of the exact build commit, commands, or Xcode/SDK installation that produced these bundles. Complete historical reproduction remains unverified.

4. License declarations on the converted artifacts

The pinned model card declares CC-BY-4.0. More specifically, Segmentation, FBank, Embedding, PLDA, and PldaRho each carry a CC-BY-4.0 license field, as do plda-parameters.json and xvector-transform.json. Thus the published declarations explicitly include the two PLDA artifacts you named. A formal maintainer confirmation of intended scope and historical build provenance would still be needed to answer those parts definitively.

The same snapshot also retains older pyannote_segmentation and wespeaker* bundles. Their metadata records a different conversion toolchain and lacks the explicit license fields above; the Community-1 provenance statements should not automatically be extended to those legacy files.