suprim/redimnet2-plus-onnx

ReDimNet2+ ONNX (FP32)

ONNX conversion of the public ReDimNet2+ LMFT speaker verification checkpoint from LAB260. The graph was exported to ONNX FP32 and validated against the original PyTorch model with ONNX Runtime CPU.

Artifact

Field	Value
File	speaker/redimnet2_plus_fp32.onnx
Input	waveform [B, T] float32, 16 kHz mono
Output	embeddings [B, 192] float32, not L2-normalized
Size	50,994,943 bytes
SHA-256	567e20568262b39b1d78fb34d0b783759ce33a8033fb6fdc0436a6bd315e17ec
Runtime	ONNX Runtime CPU; graph checked with onnx.checker
License	Apache-2.0

Provenance

• Checkpoint: lab260/redimnet2-plus, commit fd45d1968c0ea91b62ad0d2469eee4607756bb07, Apache-2.0.
• Export code: ReDimNet2plus-onnx, commit eea084c.
• Validation: speaker/redimnet2_plus_fp32.report.json; PyTorch ↔ ONNX equivalence with minimum cosine 0.99999982 and zero decision changes on the frozen evaluation set.
• Audio frontend details (peak normalization, embedding L2 norm, threshold) are deliberately outside the model and belong to the consuming application.

Usage

curl -fL -o redimnet2_plus_fp32.onnx \
  https://huggingface.co/suprim/redimnet2-plus-onnx/resolve/<commit>/speaker/redimnet2_plus_fp32.onnx
sha256sum redimnet2_plus_fp32.onnx

Download the file at a pinned commit and verify the SHA-256 before use.

Scope

Speaker verification model. It is not a surveillance system or an emotion classifier. This repository contains only the model artifact; no voice profiles or personal data are published. A change of weights may invalidate previously computed embeddings, so replace the artifact only as an explicit decision.

License

The upstream checkpoint is distributed under Apache-2.0. See LICENSE and NOTICE. The upstream training/inference code is MIT.