Add WeDefense-W2V2-Large-MHFA submission for DECRO (EER=37.7628)

submissions/wedefense-w2v2-large-mhfa.yaml

@@ -0,0 +1,54 @@
+schema_version: 4
+system:
+  name: WeDefense-W2V2-Large-MHFA
+  slug: wedefense-w2v2-large-mhfa
+  description: 'WeDefense toolkit SSL-MHFA system (Zhang et al., 2026): a wav2vec 2.0 Large (LibriSpeech-960,
+    no fine-tuning; s3prl wav2vec2_large_960) front-end fine-tuned on ASVspoof 5 together with a Multi-Head
+    Factorized Attentive Pooling back-end. All 25 hidden states (24 encoder-layer inputs + the encoder
+    output, 1024-d) are stacked, mean-normalised over the feature axis (the toolkit''s apply_cmvn) and
+    pooled by MHFA (32 heads, 128-d compression, 256-d embedding), followed by a 256->2 linear head. Trained
+    on ASVspoof 5 train, no 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 (TF32 off).'
+  code: https://github.com/zlin0/wedefense/tree/e7287fe125cefc4903ff6614afcaa6ae53842b19/egs/detection/asvspoof5/v15_ssl_mhfa
+  checkpoint: https://huggingface.co/JYP2024/Wav2vec2_Large_960-MHFA-ASV5/tree/8591bfe3eddd303af79f29b4eddad00c5d156f99
+  params_millions: 318.6973
+  paper:
+    arxiv_id: '2601.15240'
+    url: https://arxiv.org/abs/2601.15240
+    bibtex: "@article{zhang2026wedefense,\n  title={WeDefense: A Toolkit to Defend Against Fake Audio},\n\
+      \  author={Zhang, Lin and Rohdin, Johan and Wang, Xin and Peng, Junyi and Liu, Tianchi and Zhang,\
+      \ You and Luong, Hieu-Thi and Wang, Shuai and Liang, Chengdong and Silnova, Anna and Evans, Nicholas},\n\
+      \  journal={arXiv preprint arXiv:2601.15240},\n  year={2026}\n}\n"
+dataset:
+  id: SpeechAntiSpoofingBenchmarks/DECRO
+  revision: e54c29df5b81bac75e5cc294d1ab0d4db72a5c71
+  split: test
+scores:
+  eer_percent: 37.76284205472876
+  n_trials: 37314
+  n_skipped: 0
+artifact:
+  scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/WeDefense-W2V2-Large-MHFA/resolve/b0165d4b9ae2846ae5a0895c303202675b97b6fc/.eval_results/SpeechAntiSpoofingBenchmarks/DECRO/scores.txt
+  scores_sha256: 282d273cb50400cd4f6ac961472e9c562691bdae87ad2ca2fd7e294cc39f1987
+  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/Wav2vec2_Large_960-MHFA-ASV5 @ 8591bfe3eddd303af79f29b4eddad00c5d156f99, models/avg_model.pt
+  (sha256 e4cb773ea02916eecd32a24529e083bbacdddf8c4dcdd124cb98c9e208af8e17; no license on the model card;
+  linked, not re-hosted). 427 tensors, strict load, weights_only. Front-end built with fairseq 0.12.2
+  Wav2Vec2Model from the vendored model_cfg of s3prl''s converted libri960_big.pt (no Hub download; every
+  weight comes from avg_model.pt); hidden states replicate s3prl''s wav2vec2 expert hooks and S3PRLUpstream
+  length matching. MHFA, cmvn and head vendored from WeDefense @ e7287fe (Apache-2.0). Wrapper check against
+  the authors'' code (S3prlFrontend + recipe config) on ASVspoof 5 eval clips: identical logits (max |d|
+  = 0) both with the 4 s window and on whole utterances. 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 eval EER 8.08%. J-SPAW_LA scored with this wrapper (EER 2.375%).'