GoodnotesLtd/speaker-diarization-coreml-dual-output

Dual-output Core ML speaker embedding

This is a modified embedding export of FluidInference/speaker-diarization-coreml, revision 1ed7a662fdc7109e36d822db793ee6eebdaf8594, based on pyannote Community-1. Attribution belongs to the original model authors and FluidInference for the Core ML conversion. The upstream model and this derivative are distributed under CC BY 4.0. No endorsement by the upstream authors is implied.

Modifications by Goodnotes, September 2026

• Preserves the original normalized embedding output for voice identity extraction.
• Adds Float32 embedding_raw for PLDA diarization, before L2 normalization. PLDA must receive the raw vector before subtracting its trained mean.
• The added branch uses nearest/floor speaker-mask resampling from 589 to 125 frames, matching the reference ONNX embedding model. The identity branch is unchanged.
• All learned weights are unchanged. Weight SHA-256: 99356b2985b8d43880a657024d941d450b38820451ccff903f76ed4e52d1868b.

The repository contains the modified Core ML package and its compiled model. Use the original revision's FBank.mlmodelc and Segmentation.mlmodelc with this embedding model. Inputs retain the upstream model contract; this is not a standalone PCM-to-speaker pipeline. The consumer must explicitly select the appropriate output.

Verification and limitations

Export checks verify unchanged identity output on tested fixtures. A reference ONNX comparison measured raw-vector cosine similarity 0.999924 and norm ratio 1.00031. These are numerical export checks, not general speaker-recognition accuracy or identity-security guarantees. Clustering, voice enrollment, overlap handling, recording conditions, and operating thresholds require separate evaluation. No recordings, enrolled voice samples, or application code are included.

Built using coremltools 9.0 and the Apple Core ML compiler. Recompiling can change compiled artifact bytes; pin a full repository revision and verify file checksums for deployment.

Upstream research

• Plaquet and Bredin, Powerset multi-class cross entropy loss for neural speaker diarization, INTERSPEECH 2023.
• Wang et al., Wespeaker: A research and production oriented speaker embedding learning toolkit, ICASSP 2023.
• Landini et al., Bayesian HMM clustering of x-vector sequences (VBx) in speaker diarization: theory, implementation and analysis on standard tasks, Computer Speech & Language, 2022.