Add WeDefense-WavLM-Base-Pruned submission for LRLspoof (1-SRR=32.2103)

#27
by korallll - opened
submissions/wedefense-wavlm-base-pruned.yaml ADDED
@@ -0,0 +1,63 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ schema_version: 4
2
+ system:
3
+ name: WeDefense-WavLM-Base-Pruned
4
+ slug: wedefense-wavlm-base-pruned
5
+ description: 'WeDefense hybrid-pruning system (Peng et al., 2025): a WavLM Base+ front-end fine-tuned
6
+ on ASVspoof 5 jointly with structured hard-concrete pruning (target sparsity 10%: 2 attention heads
7
+ and the last FFN removed, CNN and FFN widths trimmed), then physically pruned (86.1 M params). All
8
+ 13 hidden states (CNN-projection input + 12 layers, 768-d) are stacked, mean-normalised over the feature
9
+ axis (the toolkit''s apply_cmvn on the (B, D, T, L) layout) and pooled by Multi-Head Factorized Attentive
10
+ Pooling (MHFA, 32 heads, 128-d compression, 256-d embedding), followed by BatchNorm-ReLU-Linear 256->2.
11
+ Trained on ASVspoof 5 train with MUSAN/RIR augmentation, 5 epochs, last-2 checkpoint average. Authors''
12
+ released checkpoint. Input: 16 kHz float waveform, no normalisation, first 64600 samples (4 s), shorter
13
+ clips repeat-padded (the authors score whole utterances). score = logit[bonafide] - logit[spoof].
14
+ FP32.'
15
+ code: https://github.com/zlin0/wedefense/tree/e7287fe125cefc4903ff6614afcaa6ae53842b19/egs/detection/asvspoof5/v15_ssl_mhfa_pruning
16
+ checkpoint: https://huggingface.co/JYP2024/Wedefense_ASV2025_WavLM_Base_Pruning/tree/6993dddafbbef22ca2c24bcd98ea5dab0112d9f5
17
+ params_millions: 86.11763
18
+ paper:
19
+ arxiv_id: '2508.16232'
20
+ url: https://arxiv.org/abs/2508.16232
21
+ bibtex: "@article{peng2025hybrid,\n title={Hybrid Pruning: In-Situ Compression of Self-Supervised\
22
+ \ Speech Models for Speaker Verification and Anti-Spoofing},\n author={Peng, Junyi and Zhang, Lin\
23
+ \ and Han, Jiangyu and Plchot, Old{\\v{r}}ich and Rohdin, Johan and Stafylakis, Themos and Wang,\
24
+ \ Shuai and {\\v{C}}ernock{\\'y}, Jan},\n journal={arXiv preprint arXiv:2508.16232},\n year={2025}\n\
25
+ }\n"
26
+ dataset:
27
+ id: lab260/LRLspoof
28
+ revision: ac76f0e5472fdb68a6726e1fa94a148316e76916
29
+ split: test
30
+ scores:
31
+ srr_complement: 32.210319368118704
32
+ n_trials: 1304455
33
+ n_skipped: 286
34
+ calibration:
35
+ source_dataset: SpeechAntiSpoofingBenchmarks/DeepVoice
36
+ threshold: 8.320310630573248
37
+ artifact:
38
+ scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/WeDefense-WavLM-Base-Pruned/resolve/067c4f5bd84b0396d624d9c6b6d6c4595b583559/.eval_results/lab260/LRLspoof/scores.txt
39
+ scores_sha256: 6dcba28f68cdbe5339feb04df5daf62d7d7266e043145086ef1dec5d242ce15f
40
+ bench_version: speech-spoof-bench==0.4.1
41
+ reproduction:
42
+ reproduced_by: SpeechAntiSpoofingBenchmarks
43
+ reproduced_at: '2026-11-09'
44
+ reproduced_bench_version: speech-spoof-bench==0.4.1
45
+ match: inference
46
+ submitter:
47
+ hf_username: korallll
48
+ contact: k.n.borodin@mtuci.ru
49
+ submitted_at: '2026-11-09'
50
+ notes: 'HF JYP2024/Wedefense_ASV2025_WavLM_Base_Pruning @ 6993dddafbbef22ca2c24bcd98ea5dab0112d9f5 (no
51
+ license on the model card; linked, not re-hosted). As in the authors'' recipe (run.sh stages 4-5): the
52
+ pruned front-end + MHFA come from pruned_model/whole_pytorch_model.bin (sha256 2464865501f540a65c52a017c3718ee3c9c29b9f374bd340c08d2262b2caf024),
53
+ architecture from pruned_model/pruned_config.json (vendored), and the 256->2 head from models/avg_model.pt
54
+ (sha256 a00752cbcef5cd29162409ad3959df02f290a38c938418bc612be88c7e266ed9; its other 59 non-head tensors
55
+ are the hard-concrete gates and the pruned-away layer-11 FFN, superseded by the pruned checkpoint).
56
+ Strict loads, weights_only. Code: WeDefense @ e7287fe (Apache-2.0) frontend/wav2vec2 package, MHFA.py,
57
+ projections.Linear, vendored torch-only (no s3prl / transformers). The config names wavlm_base, but
58
+ training started from wavlm-base-plus. Wrapper check: with the 4 s window disabled (whole utterance,
59
+ as the authors), 10 ASVspoof 5 eval clips reproduce the authors'' posteriors/eval/llr.txt up to the
60
+ constant log-prior offset (+2.163, std 8e-4; Spearman 1.0). Deviation: Arena 4 s cap (first 64600 samples,
61
+ repeat-pad) instead of whole utterances; the authors'' log-prior LLR shift is omitted (rank-irrelevant).
62
+ Trained on ASVspoof 5 train only (not an Arena split) -> no trained_on entries. Self-reported ASVspoof
63
+ 5 EER: dev 1.547% / eval 3.758%.'