Aivo/speaker-diarization-models

Speaker embedding model for WhisperLocal speaker detection

This repository holds the speaker embedding model that the WhisperLocal Android app downloads for on-device speaker detection (diarization). The app checks the file's size and SHA-256 after the download.

File	Size (bytes)	SHA-256
campplus_zh_en.onnx	28281164	aa3cfc16963a10586a9393f5035d6d6b57e98d358b347f80c2a30bf4f00ceba2

Origin (unmodified)

• Model: CAM++ speaker verification model iic/speech_campplus_sv_zh_en_16k-common_advanced by the 3D-Speaker project (Alibaba DAMO Academy / Tongyi Lab), https://www.modelscope.cn/models/iic/speech_campplus_sv_zh_en_16k-common_advanced and https://github.com/modelscope/3D-Speaker, licensed under Apache-2.0.
• ONNX export: k2-fsa/sherpa-onnx (Apache-2.0), release speaker-recongition-models, file 3dspeaker_speech_campplus_sv_zh_en_16k-common_advanced.onnx, https://github.com/k2-fsa/sherpa-onnx/releases/tag/speaker-recongition-models (renamed here; the file's bytes are unchanged; its ONNX metadata names the source).
• 16 kHz, 80-dimensional fbank input, 192-dimensional embeddings.

Use in WhisperLocal

Used together with pyannote segmentation 3.0 (MIT, shipped inside the app) in a C++ pipeline ported from sherpa-onnx 1.13.8. Speaker embeddings are computed on the phone, held in memory only during a run, and never stored.

License

Apache License 2.0 (see LICENSE), as the original model and the ONNX export.