Add Nes2Net-X-CtrSVDD submission for LRLspoof (1-SRR=6.0297)

#26
by korallll - opened
Files changed (1) hide show
  1. submissions/nes2net-x-ctrsvdd.yaml +61 -0
submissions/nes2net-x-ctrsvdd.yaml ADDED
@@ -0,0 +1,61 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ schema_version: 4
2
+ system:
3
+ name: Nes2Net-X-CtrSVDD
4
+ slug: nes2net-x-ctrsvdd
5
+ description: 'Singing-voice deepfake detector: Nes2Net-X trained on CtrSVDD (controlled singing voice
6
+ deepfake detection, SVDD 2024), not on speech. A fully fine-tuned WavLM-Large front-end whose 25 hidden
7
+ states are merged by squeeze-excitation layer attention (SEA), followed by the Nes2Net-X nested Res2Net
8
+ TDNN back-end (dilation 1), mean temporal pooling and a single logit (label 1 = real). Authors'' WavLM_Nes2Net_X
9
+ seed-420 epoch-75 checkpoint, FP32, deterministic first-64600-sample window (repeat-padded if shorter).
10
+ Included as an out-of-domain (singing-voice) reference.'
11
+ code: https://github.com/Liu-Tianchi/Nes2Net/tree/3fed08a603991cf8b422978e0c942a976deb9399
12
+ checkpoint: https://huggingface.co/nielsr/nes2net-checkpoints/blob/06818cabb70c44d3008fdeb9285f141f198aa92a/WavLM_Nes2Net_X_e75_seed420_valid0.03192785031473534.pt
13
+ params_millions: 315.964534
14
+ paper:
15
+ arxiv_id: '2504.05657'
16
+ url: https://arxiv.org/abs/2504.05657
17
+ bibtex: "@article{Nes2Net,\n author={Liu, Tianchi and Truong, Duc-Tuan and Das, Rohan Kumar and Lee,\
18
+ \ Kong Aik and Li, Haizhou},\n journal={IEEE Transactions on Information Forensics and Security},\n\
19
+ \ title={Nes2Net: A Lightweight Nested Architecture for Foundation Model Driven Speech Anti-Spoofing},\n\
20
+ \ year={2025},\n volume={20},\n pages={12005--12018},\n doi={10.1109/TIFS.2025.3626963}\n}\n"
21
+ dataset:
22
+ id: lab260/LRLspoof
23
+ revision: ac76f0e5472fdb68a6726e1fa94a148316e76916
24
+ split: test
25
+ scores:
26
+ srr_complement: 6.0297400106887995
27
+ n_trials: 1304455
28
+ n_skipped: 286
29
+ calibration:
30
+ source_dataset: SpeechAntiSpoofingBenchmarks/DeepVoice
31
+ threshold: 5.065291827118644
32
+ artifact:
33
+ scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/Nes2Net-X-CtrSVDD/resolve/7451ebbfa2f23940efeefca4100e8c011fbecc73/.eval_results/lab260/LRLspoof/scores.txt
34
+ scores_sha256: 4d8bb027e3c2d909c865bf755ee227161330acd413903cd0f515b30c1717dbbd
35
+ bench_version: speech-spoof-bench==0.4.1
36
+ reproduction:
37
+ reproduced_by: SpeechAntiSpoofingBenchmarks
38
+ reproduced_at: '2026-11-09'
39
+ reproduced_bench_version: speech-spoof-bench==0.4.1
40
+ match: inference
41
+ submitter:
42
+ hf_username: korallll
43
+ contact: k.n.borodin@mtuci.ru
44
+ submitted_at: '2026-11-09'
45
+ notes: 'Nes2Net-X-CtrSVDD = models/WavLM_Nes2Net_X.py (class WavLM_Nes2Net_noRes_w_allT) of Liu-Tianchi/Nes2Net
46
+ @ 3fed08a603991cf8b422978e0c942a976deb9399 (Apache-2.0) with the eval.sh settings --agg SEA --pool_func
47
+ mean --dilation 1 --Nes_ratio 8 8 --SE_ratio 1. That repository is the authors'' CtrSVDD singing-voice
48
+ repo (its README points speech users to Nes2Net_ASVspoof_ITW, which is the separate "Nes2Net" row).
49
+ Weights: WavLM_Nes2Net_X_e75_seed420_valid0.03192785031473534.pt (sha256 f1c2942635c68b584a742b10dbc5630759fc6cd55409ff8dc4d71f9c81ccdf5f)
50
+ from HF nielsr/nes2net-checkpoints @ 06818cabb70c44d3008fdeb9285f141f198aa92a (a mirror of the authors''
51
+ Google-Drive release); linked, not re-hosted (no license on the weights). The WavLM-Large architecture
52
+ is built with s3prl 0.4.18''s WavLM class from the cfg of s3prl/converted_ckpts wavlm_large.pt (vendored
53
+ as JSON, no hub download); the checkpoint overwrites every weight (strict load, 1050 tensors, 0 missing
54
+ / 0 unexpected). s3prl''s wavlm expert semantics are kept: per-utterance F.layer_norm of the waveform
55
+ (cfg normalize=True), 25 hidden states = the 24 transformer-layer inputs + the final encoder output;
56
+ verified bit-identical to the authors'' model file driven through s3prl''s UpstreamExpert on In-the-Wild
57
+ clips. Input: the authors evaluate the full utterance at batch 1 (their optional ''4s'' mode = first
58
+ 64000 samples, repeat-padded); here the Arena 4 s cap: first 64600 samples, repeat-padded if shorter,
59
+ no resampling, FP32 (cuDNN TF32 off). score = the single logit (trained with focal BCE, label 1 = bona
60
+ fide), higher = bona fide. Training data: CtrSVDD train (+ RawBoost) only, none of the Arena corpora
61
+ -> trained_on empty. Self-reported CtrSVDD test EER 2.48 % (2.22 % with ACESinger bona fide).'