CoolFace
Modelpublic

echo-hello/speech_campplus_sv_zh_en_16k-common_advanced-onnx

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face
1likes
Model Card

CAM++ Chinese-English speaker embedding (sherpa-onnx)

This repository mirrors the complete public contents of iic/speech_campplus_sv_zh_en_16k-common_advanced at ModelScope revision v1.0.0 and adds a reproducibly exported ONNX model for sherpa-onnx.

The unmodified upstream files are stored in upstream/. The original model card is preserved as upstream/README.md.

Runtime artifact

  • —File: onnx/speechcampplussvzhen16k-commonadvanced.onnx
  • —SHA-256: 945bcefe95672434d6ecf4c45bd0464d7e82be77d55985d0ec2f3253638110e1
  • —Input: float32 fbank, shape [N, T, 80]
  • —Output: float32 speaker embedding, shape [N, 192]
  • —Sample rate: 16 kHz
  • —Feature normalization: global mean
  • —Languages stated by upstream: Chinese and English

Upstream checkpoint

  • —File: upstream/camppluscnen_common.pt
  • —SHA-256: 92f29b94e6948786a26778c9e302525d185bb08c8b9f5252ed98776902840199
  • —ModelScope revision: v1.0.0

Export provenance

  • —sherpa-onnx commit: 34eba5a27220026b5981b633981c53205515067d
  • —3D-Speaker commit: 065629c313eaf1a01c65c640c46d77e61e9607b4
  • —Export script: provenance/export-onnx.py
  • —Model registry definition: provenance/infer_sv.py
  • —Export details: provenance/EXPORT.md
  • —Complete file digests: SHA256SUMS

The ONNX file is a format conversion of the upstream checkpoint. No training or fine-tuning was performed. Consumers must pin a repository commit and verify the ONNX SHA-256 instead of downloading from a floating branch.

Validation

The ONNX model passed onnx.checker.check_model. With the upstream example audio and sherpa-onnx's official test-onnx.py:

  • —same speaker: 0.6345000267028809
  • —different speaker: 0.06663094460964203

These two examples are conversion smoke tests, not a production authorization threshold. Applications must calibrate false-accept and false-reject behavior for their deployment conditions.

License and attribution

The upstream model card declares Apache License 2.0. A copy is included in LICENSE. This mirror preserves the upstream model card, attribution, examples, and configuration. The added ONNX conversion is clearly identified above.