korallll commited on
Commit
d93dfa3
·
verified ·
1 Parent(s): ac76f0e

Add xlsr-sls LRLspoof submission (srr_complement 19.68%)

Browse files
Files changed (1) hide show
  1. submissions/xlsr-sls.yaml +53 -0
submissions/xlsr-sls.yaml ADDED
@@ -0,0 +1,53 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ schema_version: 5
2
+ system:
3
+ name: XLSR-SLS
4
+ slug: xlsr-sls
5
+ description: wav2vec 2.0 (XLS-R 300M) self-supervised front-end with the SLS (Sensitive
6
+ Layer Selection) classifier for audio deepfake detection. SLS gates and fuses
7
+ the hidden states of all XLS-R transformer layers — each layer contributing distinct
8
+ discriminative cues — via a per-layer sigmoid attention, sums the weighted multi-layer
9
+ feature, then a BN + max-pool + two-layer MLP head emits a 2-way log-softmax.
10
+ Official QiShanZhang/SLSforASVspoof-2021-DF checkpoint (model_15, dev-EER 1.45%),
11
+ trained on ASVspoof2019 LA, FP32, deterministic first-64600-sample window (no
12
+ random crop).
13
+ code: https://github.com/QiShanZhang/SLSforASVspoof-2021-DF
14
+ checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/XLSR-SLS
15
+ params_millions: 340.79
16
+ paper:
17
+ arxiv_id: 10.1145/3664647.3681345
18
+ url: https://doi.org/10.1145/3664647.3681345
19
+ bibtex: "@inproceedings{zhang2024audio,\n title={Audio Deepfake Detection with\
20
+ \ Self-Supervised XLS-R and SLS Classifier},\n author={Zhang, Qishan and Wen,\
21
+ \ Shuangbing and Hu, Tao},\n booktitle={Proceedings of the 32nd ACM International\
22
+ \ Conference on Multimedia},\n pages={6765--6773},\n year={2024},\n doi={10.1145/3664647.3681345}\n\
23
+ }\n"
24
+ dataset:
25
+ id: lab260/LRLspoof
26
+ revision: 5caf11349cf96aa198527cfb040cdba585cdd49e
27
+ split: test
28
+ scores:
29
+ srr_complement: 19.675210804734668
30
+ n_trials: 1304455
31
+ n_skipped: 286
32
+ calibration:
33
+ source_dataset: SpeechAntiSpoofingBenchmarks/DeepVoice
34
+ threshold: -4.67828299044586
35
+ artifact:
36
+ scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/XLSR-SLS/resolve/f67c691393171b9916d2e44c00c987bd13afebd1/.eval_results/lab260/LRLspoof/scores.txt
37
+ scores_sha256: c0b68dead9f8a31054100608c8f9763855cb2b433532c0ed60c7d9f008649304
38
+ bench_version: speech-spoof-bench==0.4.1
39
+ reproduction:
40
+ reproduced_by: SpeechAntiSpoofingBenchmarks
41
+ reproduced_at: '2026-06-22'
42
+ reproduced_bench_version: speech-spoof-bench==0.4.1
43
+ match: scoring
44
+ submitter:
45
+ hf_username: korallll
46
+ contact: k.n.borodin@mtuci.ru
47
+ submitted_at: '2026-06-22'
48
+ notes: XLSR-SLS on LRLspoof (spoof-only, srr_complement = 1-SRR, lower better). TensorRT
49
+ FP32 engine (parity-verified vs PyTorch). Scored 1,304,169 of 1,304,455 TTS utterances;
50
+ 286 source files are empty/zero-length (skipped), 172 corrupt-header WAVs recovered
51
+ via ffmpeg. Audio resampled per-file to 16 kHz; score = net(x)[:,1] (higher = bona
52
+ fide). t* = this model's own DeepVoice EER operating-point threshold (DeepVoice
53
+ EER 27.07%), transferred via the calibration block. Verified with reproduce --scoring.