FluidInference/speaker-diarization-coreml

Speaker Diarization Core ML

Core ML conversions used by FluidAudio for on-device speaker diarization on Apple platforms.

The supported Community-1 artifact set is distributed under CC-BY-4.0. See NOTICE.md for attribution and the precise license scope, and PROVENANCE.md plus provenance.json for source, conversion, environment, and file-integrity records.

Supported Community-1 artifacts

Artifact	Source
Segmentation.mlmodelc	pyannote/speaker-diarization-community-1/segmentation/pytorch_model.bin
FBank.mlmodelc	Deterministic feature-extraction graph configured from the embedding checkpoint
Embedding.mlmodelc	pyannote/speaker-diarization-community-1/embedding/pytorch_model.bin
PLDA.mlmodelc, PldaRho.mlmodelc	plda/plda.npz and plda/xvec_transform.npz
plda-parameters.json, xvector-transform.json	Serialized tensors from the same two PLDA files

The snapshot includes uncompiled packages for Segmentation, FBank, Embedding, and PldaRho under mlpackages/; no uncompiled PLDA package was published. These are Core ML conversions, not fine-tuned models.

Provenance status

The artifacts are pinned by SHA-256 in provenance.json. Their historical source lineage has been reconstructed from the public Community-1 repository, artifact metadata, and the published conversion source. The exact local upstream checkout and Mobius commit used for the original 2025 conversion were not recorded, so the existing binaries are not described as a fully attested reproducible build.

The immutable upstream reference used for reconstruction is pyannote/speaker-diarization-community-1@3533c8cf. The public historical conversion reference is FluidInference/mobius@33fd6eab.

The current conversion pipeline requires a full upstream commit SHA, records input and output hashes, embeds source metadata in each model, and captures the build environment. It is available at FluidInference/mobius@ffbc3c8.

Legacy compatibility artifacts

The repository also retains pyannote_segmentation.mlmodelc, wespeaker.mlmodelc, wespeaker_v2.mlmodelc, and wespeaker_int8.mlmodelc for FluidAudio's legacy online diarizer. They predate the Community-1 export, record a different toolchain, and are not included in the Community-1 provenance or license-scope confirmation in NOTICE.md.

Technical specifications

• Input: 16 kHz mono audio
• Output: speaker segments with timestamps and speaker identifiers
• Framework: Core ML converted from PyTorch
• Deployment target: iOS 17 / macOS 14 or later
• Community-1 conversion metadata: PyTorch 2.8.0, coremltools 9.0b1, TorchScript source dialect
• Compiled MIL metadata: coremlc 3500.32.1, MIL 3500.14.1

Usage

See the FluidAudio diarization documentation.

Citations

Speaker segmentation

@inproceedings{Plaquet23,
  author={Alexis Plaquet and Hervé Bredin},
  title={{Powerset multi-class cross entropy loss for neural speaker diarization}},
  year=2023,
  booktitle={Proc. INTERSPEECH 2023},
}

Speaker embedding

@inproceedings{Wang2023,
  title={Wespeaker: A research and production oriented speaker embedding learning toolkit},
  author={Wang, Hongji and Liang, Chengdong and Wang, Shuai and Chen, Zhengyang and Zhang, Binbin and Xiang, Xu and Deng, Yanlei and Qian, Yanmin},
  booktitle={ICASSP 2023, IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages={1--5},
  year={2023},
  organization={IEEE}
}

Speaker clustering

@article{Landini2022,
  author={Landini, Federico and Profant, J{\'a}n and Diez, Mireia and Burget, Luk{\'a}{\v{s}}},
  title={{Bayesian HMM clustering of x-vector sequences (VBx) in speaker diarization: theory, implementation and analysis on standard tasks}},
  year={2022},
  journal={Computer Speech & Language},
}