Add Nes2Net-X-CtrSVDD submission for DECRO (EER=35.9344)

submissions/nes2net-x-ctrsvdd.yaml

@@ -0,0 +1,58 @@
+schema_version: 4
+system:
+  name: Nes2Net-X-CtrSVDD
+  slug: nes2net-x-ctrsvdd
+  description: 'Singing-voice deepfake detector: Nes2Net-X trained on CtrSVDD (controlled singing voice
+    deepfake detection, SVDD 2024), not on speech. A fully fine-tuned WavLM-Large front-end whose 25 hidden
+    states are merged by squeeze-excitation layer attention (SEA), followed by the Nes2Net-X nested Res2Net
+    TDNN back-end (dilation 1), mean temporal pooling and a single logit (label 1 = real). Authors'' WavLM_Nes2Net_X
+    seed-420 epoch-75 checkpoint, FP32, deterministic first-64600-sample window (repeat-padded if shorter).
+    Included as an out-of-domain (singing-voice) reference.'
+  code: https://github.com/Liu-Tianchi/Nes2Net/tree/3fed08a603991cf8b422978e0c942a976deb9399
+  checkpoint: https://huggingface.co/nielsr/nes2net-checkpoints/blob/06818cabb70c44d3008fdeb9285f141f198aa92a/WavLM_Nes2Net_X_e75_seed420_valid0.03192785031473534.pt
+  params_millions: 315.964534
+  paper:
+    arxiv_id: '2504.05657'
+    url: https://arxiv.org/abs/2504.05657
+    bibtex: "@article{Nes2Net,\n  author={Liu, Tianchi and Truong, Duc-Tuan and Das, Rohan Kumar and Lee,\
+      \ Kong Aik and Li, Haizhou},\n  journal={IEEE Transactions on Information Forensics and Security},\n\
+      \  title={Nes2Net: A Lightweight Nested Architecture for Foundation Model Driven Speech Anti-Spoofing},\n\
+      \  year={2025},\n  volume={20},\n  pages={12005--12018},\n  doi={10.1109/TIFS.2025.3626963}\n}\n"
+dataset:
+  id: SpeechAntiSpoofingBenchmarks/DECRO
+  revision: e54c29df5b81bac75e5cc294d1ab0d4db72a5c71
+  split: test
+scores:
+  eer_percent: 35.934421353953674
+  n_trials: 37314
+  n_skipped: 0
+artifact:
+  scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/Nes2Net-X-CtrSVDD/resolve/acaab5409c657f05a65e9edc492fa0fcb3562353/.eval_results/SpeechAntiSpoofingBenchmarks/DECRO/scores.txt
+  scores_sha256: 388d9a7e5f592a81fdbf40f2e38927bd568124cd71946eb57ed17f855342c630
+  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: 'Nes2Net-X-CtrSVDD = models/WavLM_Nes2Net_X.py (class WavLM_Nes2Net_noRes_w_allT) of Liu-Tianchi/Nes2Net
+  @ 3fed08a603991cf8b422978e0c942a976deb9399 (Apache-2.0) with the eval.sh settings --agg SEA --pool_func
+  mean --dilation 1 --Nes_ratio 8 8 --SE_ratio 1. That repository is the authors'' CtrSVDD singing-voice
+  repo (its README points speech users to Nes2Net_ASVspoof_ITW, which is the separate "Nes2Net" row).
+  Weights: WavLM_Nes2Net_X_e75_seed420_valid0.03192785031473534.pt (sha256 f1c2942635c68b584a742b10dbc5630759fc6cd55409ff8dc4d71f9c81ccdf5f)
+  from HF nielsr/nes2net-checkpoints @ 06818cabb70c44d3008fdeb9285f141f198aa92a (a mirror of the authors''
+  Google-Drive release); linked, not re-hosted (no license on the weights). The WavLM-Large architecture
+  is built with s3prl 0.4.18''s WavLM class from the cfg of s3prl/converted_ckpts wavlm_large.pt (vendored
+  as JSON, no hub download); the checkpoint overwrites every weight (strict load, 1050 tensors, 0 missing
+  / 0 unexpected). s3prl''s wavlm expert semantics are kept: per-utterance F.layer_norm of the waveform
+  (cfg normalize=True), 25 hidden states = the 24 transformer-layer inputs + the final encoder output;
+  verified bit-identical to the authors'' model file driven through s3prl''s UpstreamExpert on In-the-Wild
+  clips. Input: the authors evaluate the full utterance at batch 1 (their optional ''4s'' mode = first
+  64000 samples, repeat-padded); here the Arena 4 s cap: first 64600 samples, repeat-padded if shorter,
+  no resampling, FP32 (cuDNN TF32 off). score = the single logit (trained with focal BCE, label 1 = bona
+  fide), higher = bona fide. Training data: CtrSVDD train (+ RawBoost) only, none of the Arena corpora
+  -> trained_on empty. Self-reported CtrSVDD test EER 2.48 % (2.22 % with ACESinger bona fide).'