Add WeDefense-WavLM-Base-Pruned submission for DECRO (EER=40.5777)

submissions/wedefense-wavlm-base-pruned.yaml

@@ -0,0 +1,60 @@
+schema_version: 4
+system:
+  name: WeDefense-WavLM-Base-Pruned
+  slug: wedefense-wavlm-base-pruned
+  description: '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.'
+  code: https://github.com/zlin0/wedefense/tree/e7287fe125cefc4903ff6614afcaa6ae53842b19/egs/detection/asvspoof5/v15_ssl_mhfa_pruning
+  checkpoint: https://huggingface.co/JYP2024/Wedefense_ASV2025_WavLM_Base_Pruning/tree/6993dddafbbef22ca2c24bcd98ea5dab0112d9f5
+  params_millions: 86.11763
+  paper:
+    arxiv_id: '2508.16232'
+    url: https://arxiv.org/abs/2508.16232
+    bibtex: "@article{peng2025hybrid,\n  title={Hybrid Pruning: In-Situ Compression of Self-Supervised\
+      \ Speech Models for Speaker Verification and Anti-Spoofing},\n  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},\n  journal={arXiv preprint arXiv:2508.16232},\n  year={2025}\n\
+      }\n"
+dataset:
+  id: SpeechAntiSpoofingBenchmarks/DECRO
+  revision: e54c29df5b81bac75e5cc294d1ab0d4db72a5c71
+  split: test
+scores:
+  eer_percent: 40.577716643741404
+  n_trials: 37314
+  n_skipped: 0
+artifact:
+  scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/WeDefense-WavLM-Base-Pruned/resolve/4b312ea1ebf5114b28153d52bb1e35c3ca92df2c/.eval_results/SpeechAntiSpoofingBenchmarks/DECRO/scores.txt
+  scores_sha256: 84f4935efeb2f7699b5c2f4dab16da30a7c3bf4b5da27d0768cf2cadc7fd0c75
+  bench_version: speech-spoof-bench==0.4.1
+reproduction:
+  reproduced_by: SpeechAntiSpoofingBenchmarks
+  reproduced_at: '2026-11-08'
+  reproduced_bench_version: speech-spoof-bench==0.4.1
+  match: inference
+submitter:
+  hf_username: korallll
+  contact: k.n.borodin@mtuci.ru
+submitted_at: '2026-11-08'
+notes: 'HF JYP2024/Wedefense_ASV2025_WavLM_Base_Pruning @ 6993dddafbbef22ca2c24bcd98ea5dab0112d9f5 (no
+  license on the model card; linked, not re-hosted). As in the authors'' recipe (run.sh stages 4-5): the
+  pruned front-end + MHFA come from pruned_model/whole_pytorch_model.bin (sha256 2464865501f540a65c52a017c3718ee3c9c29b9f374bd340c08d2262b2caf024),
+  architecture from pruned_model/pruned_config.json (vendored), and the 256->2 head from models/avg_model.pt
+  (sha256 a00752cbcef5cd29162409ad3959df02f290a38c938418bc612be88c7e266ed9; its other 59 non-head tensors
+  are the hard-concrete gates and the pruned-away layer-11 FFN, superseded by the pruned checkpoint).
+  Strict loads, weights_only. Code: WeDefense @ e7287fe (Apache-2.0) frontend/wav2vec2 package, MHFA.py,
+  projections.Linear, vendored torch-only (no s3prl / transformers). The config names wavlm_base, but
+  training started from wavlm-base-plus. Wrapper check: with the 4 s window disabled (whole utterance,
+  as the authors), 10 ASVspoof 5 eval clips reproduce the authors'' posteriors/eval/llr.txt up to the
+  constant log-prior offset (+2.163, std 8e-4; Spearman 1.0). Deviation: Arena 4 s cap (first 64600 samples,
+  repeat-pad) instead of whole utterances; the authors'' log-prior LLR shift is omitted (rank-irrelevant).
+  Trained on ASVspoof 5 train only (not an Arena split) -> no trained_on entries. Self-reported ASVspoof
+  5 EER: dev 1.547% / eval 3.758%.'