Add XLSR-Nes2Net-LA submission for LRLspoof (1-SRR=0.7240)

#23
by korallll - opened
Files changed (1) hide show
  1. submissions/xlsr-nes2net-la.yaml +62 -0
submissions/xlsr-nes2net-la.yaml ADDED
@@ -0,0 +1,62 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ schema_version: 4
2
+ system:
3
+ name: XLSR-Nes2Net-LA
4
+ slug: xlsr-nes2net-la
5
+ description: 'XLSR-Nes2Net-LA (Zhang et al., MultiAPI Spoof, 2025): a wav2vec 2.0 XLS-R 300M front-end
6
+ fine-tuned end-to-end with a Nes2Net-X (nested Res2Net TDNN) back-end extended by a local-window cross-block
7
+ attention (window 3, single head, learned block position embedding) between the nested blocks, followed
8
+ by mean temporal pooling and a 2-way linear classifier. Trained on a multi-corpus mix (TIMIT, ODSS,
9
+ FoR, AI4T, ASVspoof 5, MLAAD and MultiAPI-Spoof). Authors'' released epoch_5_1.422.pth checkpoint,
10
+ FP32, deterministic first-64600-sample window (repeat-padded if shorter), no normalisation. score
11
+ = logit of the bona fide class.'
12
+ code: https://github.com/XuepingZhang/MultiAPI-Spoof/tree/1f712e3fff26e237268385b92365a9debd4cb69f
13
+ checkpoint: https://huggingface.co/XuepingZhang/XLSR-Nes2Net-LA/blob/ef20423b7a2555f699f217184f3629078bf0e664/epoch_5_1.422.pth
14
+ params_millions: 317.969191
15
+ paper:
16
+ arxiv_id: '2512.07352'
17
+ url: https://arxiv.org/abs/2512.07352
18
+ bibtex: "@misc{zhang2025multiapispoofmultiapidataset,\n title={MultiAPI Spoof: A Multi-API Dataset\
19
+ \ and Local-Attention Network for Speech Anti-spoofing Detection},\n author={Xueping Zhang\
20
+ \ and Zhenshan Zhang and Yechen Wang and Linxi Li and Liwei Jin and Ming Li},\n year={2025},\n\
21
+ \ eprint={2512.07352},\n archivePrefix={arXiv},\n primaryClass={cs.SD},\n url={https://arxiv.org/abs/2512.07352},\n\
22
+ }\n"
23
+ dataset:
24
+ id: lab260/LRLspoof
25
+ revision: ac76f0e5472fdb68a6726e1fa94a148316e76916
26
+ split: test
27
+ scores:
28
+ srr_complement: 0.7239859251370051
29
+ n_trials: 1304455
30
+ n_skipped: 286
31
+ calibration:
32
+ source_dataset: SpeechAntiSpoofingBenchmarks/DeepVoice
33
+ threshold: 0.7138218789808917
34
+ artifact:
35
+ scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/XLSR-Nes2Net-LA/resolve/75e2b1b0b1a2ab5816c45fceb41b3975726323f9/.eval_results/lab260/LRLspoof/scores.txt
36
+ scores_sha256: 92fafa9d0da37c8674ee00d4b43ffbf9b58d974b7bdf8d2c61736ec884ea59cc
37
+ bench_version: speech-spoof-bench==0.4.1
38
+ reproduction:
39
+ reproduced_by: SpeechAntiSpoofingBenchmarks
40
+ reproduced_at: '2026-11-09'
41
+ reproduced_bench_version: speech-spoof-bench==0.4.1
42
+ match: inference
43
+ submitter:
44
+ hf_username: korallll
45
+ contact: k.n.borodin@mtuci.ru
46
+ submitted_at: '2026-11-09'
47
+ notes: 'XLSR-Nes2Net-LA = models/XLSR2_Nes2Net_X_ATT.py of XuepingZhang/MultiAPI-Spoof @ 1f712e3fff26e237268385b92365a9debd4cb69f
48
+ (code CC BY-NC 4.0) with job/xlsr2_nes2net_ATT/few_dkucc.conf (Nes_ratio [8,8], SE_ratio [1], dilation
49
+ 2, pool_func mean, 2 logits). Weights: HF XuepingZhang/XLSR-Nes2Net-LA @ ef20423b7a2555f699f217184f3629078bf0e664,
50
+ epoch_5_1.422.pth (sha256 c855f654f640b39fa19295445d1c67dc3658060f2282e015208eaedb5a34ddf8), linked,
51
+ not re-hosted (the weights carry no license). The XLS-R 300M architecture is built from fairseq xlsr2_300m.pt
52
+ and every weight is overwritten by the checkpoint (strict load after stripping the DataParallel `module.`
53
+ prefix; 998 tensors, 0 missing / 0 unexpected). The source''s `global_attn` (nn.MultiheadAttention(1024,
54
+ 8), 4.2M params) is stored in the checkpoint but never used by forward(); it is constructed only so
55
+ the load is strict, and params_millions counts the forward path (322.17M incl. it). Input: the authors
56
+ train and evaluate on random 64600-sample crops (pad_random); here a deterministic first-64600-sample
57
+ window, repeat-padded if shorter, no normalisation, no resampling, FP32 (cuDNN TF32 off). score = logits[:,
58
+ 1] (bona fide = class 1, as the source''s evaluation), higher = bona fide. Training data (paper / repo):
59
+ TIMIT, ODSS, FoR, AI4T, ASVspoof 5 (train), MLAAD and MultiAPI-Spoof train -> ODSS is marked trained_on.
60
+ In-the-Wild was used as the development set for checkpoint selection (the file name is the ITW EER 1.422
61
+ %); by maintainer decision ITW stays ranked and is not marked trained_on. Self-reported EER: ITW 1.42
62
+ %, MultiAPI-Spoof 0.56 %, AI4T 5.64 %.'