korallll commited on
Commit
0946ab0
·
1 Parent(s): 7935dc8

Add RawNetLite submission for LRLspoof (1-SRR=10.4843) (#41)

Browse files

- Add RawNetLite submission for LRLspoof (1-SRR=10.4843) (81654ee9af139d7d83c6e09f9f34addad9233472)
- Fill reproduction block (inference, maintainer re-ran the model) (20a745e477aaf8799736fdbb817d995ff22249fd)

Files changed (1) hide show
  1. submissions/rawnetlite.yaml +64 -0
submissions/rawnetlite.yaml ADDED
@@ -0,0 +1,64 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ schema_version: 4
2
+ system:
3
+ name: RawNetLite
4
+ slug: rawnetlite
5
+ description: 'RawNetLite (Di Pierno et al., IJCNN VERIMEDIA 2025): a lightweight end-to-end raw-waveform
6
+ detector -- Conv1D(1->64) + BatchNorm + ReLU, three 1-D residual blocks (64 ch), adaptive average
7
+ pooling to 64 steps, a bidirectional GRU (128 per direction, last step), Linear(256,64) -> Linear(64,1).
8
+ Released checkpoint triple_cross_domain_focal_rawnet_lite.pt: trained with focal loss on a balanced
9
+ mix of FakeOrReal, "AVSpoof2021" (the authors'' name for ASVspoof 2021) and CodecFake, no augmentation.
10
+ Input as the authors'' audio_preprocessor.py: 16 kHz mono, peak-normalised (divide by max |x| of the
11
+ whole clip), first 48000 samples (3 s, within the Arena 4 s cap), right zero-pad. score = -logit(fake)
12
+ (the pre-sigmoid fc2 output negated; the source''s sigmoid(logit) is P(fake)). FP32 (TF32 off).'
13
+ code: https://github.com/adipiz99/RawNetLite/tree/fcb0fe8d0f40146d110e322c00a0723ba27d5520
14
+ checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/RawNetLite/blob/7209f7f94f1c69bc0d2320e854914e0966b12bd4/triple_cross_domain_focal_rawnet_lite.pt
15
+ params_millions: 0.240769
16
+ paper:
17
+ arxiv_id: '2504.20923'
18
+ url: https://arxiv.org/abs/2504.20923
19
+ bibtex: "@inproceedings{dipierno2025rawnetlite,\n title = {End-to-end Audio Deepfake Detection\
20
+ \ from RAW Waveforms: a RawNet-Based Approach with Cross-Dataset Evaluation},\n author = {Andrea\
21
+ \ Di Pierno and Luca Guarnera and Dario Allegra and Sebastiano Battiato},\n booktitle = {International\
22
+ \ Joint Conference on Neural Networks (IJCNN) - VERIMEDIA Workshop},\n year = {2025}\n}\n"
23
+ dataset:
24
+ id: lab260/LRLspoof
25
+ revision: ac76f0e5472fdb68a6726e1fa94a148316e76916
26
+ split: test
27
+ scores:
28
+ srr_complement: 10.484300730963547
29
+ n_trials: 1304455
30
+ n_skipped: 286
31
+ calibration:
32
+ source_dataset: SpeechAntiSpoofingBenchmarks/DeepVoice
33
+ threshold: 1.586388167197452
34
+ artifact:
35
+ scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/RawNetLite/resolve/4f73e46bc17a437b1f20ce466887b840da3364bf/.eval_results/lab260/LRLspoof/scores.txt
36
+ scores_sha256: af4317b53cae055bf06099a8b7a6b6fd0dcaee8a6a0c534d68e3faad2ffd6e18
37
+ bench_version: speech-spoof-bench==0.4.1
38
+ reproduction:
39
+ reproduced_by: SpeechAntiSpoofingBenchmarks
40
+ reproduced_at: '2026-11-16'
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: '2026-11-16'
47
+ notes: 'RawNetLite = github.com/adipiz99/RawNetLite @ fcb0fe8d0f40146d110e322c00a0723ba27d5520 (MIT; RawNetLite.py
48
+ and audio_preprocessor.py vendored), checkpoint models/triple_cross_domain_focal_rawnet_lite.pt (sha256
49
+ 6154d91f34f0a49d58e30bac602915f6b6853419419804be63b5274603891a9b; plain state_dict, 61 tensors, weights_only,
50
+ strict load, all used; re-hosted under MIT). Wrapper vs the authors'' RawNetLite + preprocess_audio
51
+ + tester.py scoring on J-SPAW clips: max |d P(fake)| 8.5e-8. Deviations: score is -logit instead of
52
+ the sigmoid output (same ranking, avoids float saturation ties); a silent clip (source: division by
53
+ zero -> NaN) is scored as zeros; no resampling (authors'' target is 16 kHz too). The authors'' 3 s window
54
+ is kept (shorter than the 4 s cap). Training data (paper Tables I-III + README): FakeOrReal (25,600+25,600),
55
+ "AVSpoof2021" 6,400+6,400 and CodecFake 6,400+6,400 training clips. The paper never names the ASVspoof
56
+ 2021 track and 2021 has no train split (the training clips come from the released evaluation data),
57
+ so both ASVspoof2021_LA and ASVspoof2021_DF are marked trained_on (n/a*). Its citation for "AVSpoof2021"
58
+ is the ASVspoof 2019 overview paper and it mentions replay attacks, so ASVspoof 2019 LA/PA or 2021 PA
59
+ data cannot be excluded either (not marked). CodecFake is linked to roger-tseng/CodecFake (Wu et al.,
60
+ VCTK-based) in the README and cited as Xie et al.''s Codecfake in the paper; neither is the Arena''s
61
+ CD-ADD (Li et al., LibriTTS-based), and FakeOrReal is not an Arena set -> no further entries. Self-reported
62
+ EER% (paper Table XII, this checkpoint): FoR 2.66 / AVSpoof2021 19.96 / CodecFake 21.13 / triple-domain
63
+ test 16.44. 2000-clip eval subsets with this wrapper: ASVspoof2019_LA 13.82, ASVspoof2021_LA 20.24,
64
+ ASVspoof2021_DF 21.25. J-SPAW_LA scored with this wrapper (EER 34.19%).'