Provenance and licence of the legacy pyannote_segmentation / wespeaker_v2 diarizer models #1010

Hello, and thanks for FluidAudio.

We use FluidAudio (at 20d4f0bd) in a paid macOS app for dictation and meeting transcription. For speaker voiceprints we call DiarizerModels.download() and DiarizerManager.extractSpeakerEmbedding(from:), which fetch and run pyannote_segmentation.mlmodelc and wespeaker_v2.mlmodelc from FluidInference/speaker-diarization-coreml. The models are downloaded at runtime from Hugging Face; we do not ship them in our installer.

The model repository's NOTICE.md and README.md say these legacy files are outside the Community-1 licence scope and that their source and licensing must be evaluated separately. #927 covered the Community-1 artifacts; this is the same question for the legacy ones, which are still what DiarizerModels downloads by default.

Could you clarify:

1. Upstream checkpoints. Which checkpoints (name and revision) were converted into pyannote_segmentation.mlmodelc and wespeaker_v2.mlmodelc? Our guess is pyannote/segmentation-3.0 (MIT) and pyannote/wespeaker-voxceleb-resnet34-LM (CC-BY-4.0). The metadata only records the toolchain (torch 2.6.0, coremltools 8.3.0), and the generated class name of wespeaker_v2 is wespeaker_int8, so we cannot tell which checkpoint and quantisation it comes from.
2. Licence. Under what licence does Fluid Inference publish these two converted files, and may a commercial app download and run them?
3. Way forward. If the legacy pair cannot be cleared, do you plan to move DiarizerModels to the Community-1 Embedding + FBank artifacts? Is there, or will there be, an API to get a single speaker embedding from those, as extractSpeakerEmbedding(from:) does today? We assume their vectors are not compatible with those from wespeaker_v2, so our users would have to rebuild their saved voiceprints. Please correct us if that is wrong.

Thank you.