Add XLSR-Nes2Net-LA submission for DECRO (EER=2.7748)

submissions/xlsr-nes2net-la.yaml

@@ -0,0 +1,59 @@
+schema_version: 4
+system:
+  name: XLSR-Nes2Net-LA
+  slug: xlsr-nes2net-la
+  description: 'XLSR-Nes2Net-LA (Zhang et al., MultiAPI Spoof, 2025): a wav2vec 2.0 XLS-R 300M front-end
+    fine-tuned end-to-end with a Nes2Net-X (nested Res2Net TDNN) back-end extended by a local-window cross-block
+    attention (window 3, single head, learned block position embedding) between the nested blocks, followed
+    by mean temporal pooling and a 2-way linear classifier. Trained on a multi-corpus mix (TIMIT, ODSS,
+    FoR, AI4T, ASVspoof 5, MLAAD and MultiAPI-Spoof). Authors'' released epoch_5_1.422.pth checkpoint,
+    FP32, deterministic first-64600-sample window (repeat-padded if shorter), no normalisation. score
+    = logit of the bona fide class.'
+  code: https://github.com/XuepingZhang/MultiAPI-Spoof/tree/1f712e3fff26e237268385b92365a9debd4cb69f
+  checkpoint: https://huggingface.co/XuepingZhang/XLSR-Nes2Net-LA/blob/ef20423b7a2555f699f217184f3629078bf0e664/epoch_5_1.422.pth
+  params_millions: 317.969191
+  paper:
+    arxiv_id: '2512.07352'
+    url: https://arxiv.org/abs/2512.07352
+    bibtex: "@misc{zhang2025multiapispoofmultiapidataset,\n      title={MultiAPI Spoof: A Multi-API Dataset\
+      \ and Local-Attention Network for Speech Anti-spoofing Detection},\n      author={Xueping Zhang\
+      \ and Zhenshan Zhang and Yechen Wang and Linxi Li and Liwei Jin and Ming Li},\n      year={2025},\n\
+      \      eprint={2512.07352},\n      archivePrefix={arXiv},\n      primaryClass={cs.SD},\n      url={https://arxiv.org/abs/2512.07352},\n\
+      }\n"
+dataset:
+  id: SpeechAntiSpoofingBenchmarks/DECRO
+  revision: e54c29df5b81bac75e5cc294d1ab0d4db72a5c71
+  split: test
+scores:
+  eer_percent: 2.774843975036006
+  n_trials: 37314
+  n_skipped: 0
+artifact:
+  scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/XLSR-Nes2Net-LA/resolve/1a1f042fe7ee75ae2bf77d4d0a686d2957eba022/.eval_results/SpeechAntiSpoofingBenchmarks/DECRO/scores.txt
+  scores_sha256: c5982b84b1e0a098352769c1ebf47077302ecac9b971de48eae7579d16653382
+  bench_version: speech-spoof-bench==0.4.1
+reproduction:
+  reproduced_by: SpeechAntiSpoofingBenchmarks
+  reproduced_at: '2026-11-07'
+  reproduced_bench_version: speech-spoof-bench==0.4.1
+  match: inference
+submitter:
+  hf_username: korallll
+  contact: k.n.borodin@mtuci.ru
+submitted_at: '2026-11-07'
+notes: 'XLSR-Nes2Net-LA = models/XLSR2_Nes2Net_X_ATT.py of XuepingZhang/MultiAPI-Spoof @ 1f712e3fff26e237268385b92365a9debd4cb69f
+  (code CC BY-NC 4.0) with job/xlsr2_nes2net_ATT/few_dkucc.conf (Nes_ratio [8,8], SE_ratio [1], dilation
+  2, pool_func mean, 2 logits). Weights: HF XuepingZhang/XLSR-Nes2Net-LA @ ef20423b7a2555f699f217184f3629078bf0e664,
+  epoch_5_1.422.pth (sha256 c855f654f640b39fa19295445d1c67dc3658060f2282e015208eaedb5a34ddf8), linked,
+  not re-hosted (the weights carry no license). The XLS-R 300M architecture is built from fairseq xlsr2_300m.pt
+  and every weight is overwritten by the checkpoint (strict load after stripping the DataParallel `module.`
+  prefix; 998 tensors, 0 missing / 0 unexpected). The source''s `global_attn` (nn.MultiheadAttention(1024,
+  8), 4.2M params) is stored in the checkpoint but never used by forward(); it is constructed only so
+  the load is strict, and params_millions counts the forward path (322.17M incl. it). Input: the authors
+  train and evaluate on random 64600-sample crops (pad_random); here a deterministic first-64600-sample
+  window, repeat-padded if shorter, no normalisation, no resampling, FP32 (cuDNN TF32 off). score = logits[:,
+  1] (bona fide = class 1, as the source''s evaluation), higher = bona fide. Training data (paper / repo):
+  TIMIT, ODSS, FoR, AI4T, ASVspoof 5 (train), MLAAD and MultiAPI-Spoof train -> ODSS is marked trained_on.
+  In-the-Wild was used as the development set for checkpoint selection (the file name is the ITW EER 1.422
+  %); by maintainer decision ITW stays ranked and is not marked trained_on. Self-reported EER: ITW 1.42
+  %, MultiAPI-Spoof 0.56 %, AI4T 5.64 %.'