WeDefense-WavLM-Base-Pruned: Speech Anti-Spoofing Arena wrapper

Packaged for the Speech Anti-Spoofing Arena from the authors' code and paper.

WeDefense hybrid-pruning system (Peng et al., 2025): a WavLM Base+ front-end fine-tuned on ASVspoof 5 jointly with structured hard-concrete pruning (target sparsity 10%: 2 attention heads and the last FFN removed, CNN and FFN widths trimmed), then physically pruned (86.1 M params). All 13 hidden states (CNN-projection input + 12 layers, 768-d) are stacked, mean-normalised over the feature axis (the toolkit's apply_cmvn on the (B, D, T, L) layout) and pooled by Multi-Head Factorized Attentive Pooling (MHFA, 32 heads, 128-d compression, 256-d embedding), followed by BatchNorm-ReLU-Linear 256->2. Trained on ASVspoof 5 train with MUSAN/RIR augmentation, 5 epochs, last-2 checkpoint average. Authors' released checkpoint. Input: 16 kHz float waveform, no normalisation, first 64600 samples (4 s), shorter clips repeat-padded (the authors score whole utterances). score = logit[bonafide] - logit[spoof]. FP32.

• Weights: not re-hosted (the authors grant no license for them). Get the authors' checkpoint; its sha256 is in the meta.yaml notes.
• Params: 86.11763 M. Trained on ASVspoof 5 train with MUSAN/RIR augmentation (not an Arena split).
• Score: higher = bona fide. J-SPAW_LA was scored with the PyTorch wrapper; the other datasets with the TensorRT drop-in class (FP32 on the RTX 4070 Ti SUPER; FP32 on the RTX 5060 Ti), which matched PyTorch on the full J-SPAW_LA set.

Arena results

EER% (lower is better); EmoSpoofTTS and LRLspoof are spoof-only and use 1-SRR% at the model's DeepVoice EER threshold. Per-dataset scores are under .eval_results/.

Citation

@article{peng2025hybrid,
  title={Hybrid Pruning: In-Situ Compression of Self-Supervised Speech Models for Speaker Verification and Anti-Spoofing},
  author={Peng, Junyi and Zhang, Lin and Han, Jiangyu and Plchot, Old{\v{r}}ich and Rohdin, Johan and Stafylakis, Themos and Wang, Shuai and {\v{C}}ernock{\'y}, Jan},
  journal={arXiv preprint arXiv:2508.16232},
  year={2025}
}