korallll commited on
Commit
40e2a5e
·
verified ·
1 Parent(s): b35eb9c

Add SpoofSUPERB-UniSpeechSAT submission for CFAD (EER=16.4436)

Browse files
submissions/spoofsuperb-unispeechsat.yaml ADDED
@@ -0,0 +1,52 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ schema_version: 4
2
+ system:
3
+ name: SpoofSUPERB-UniSpeechSAT
4
+ slug: spoofsuperb-unispeechsat
5
+ description: 'Spoof-SUPERB linear-head detector (Ali et al., 2026) over a frozen UniSpeech-SAT Large
6
+ (24 layers, 1024-d; s3prl `unispeech_sat_large`): the s3prl Featurizer softmax-weights all 25 hidden
7
+ states (24 transformer-layer inputs + the encoder output; no per-layer norm), then Linear(1024,256)
8
+ -> mean pooling over frames -> Linear(256,2). Only the layer weights, projector and classifier are
9
+ trained (ASVspoof 2019 LA train, 50 epochs, SWA); the authors'' released head with the s3prl converted
10
+ upstream. Input as the authors'' scoring code: 16 kHz, first 64600 samples (shorter clips tiled),
11
+ the upstream''s own per-utterance waveform layer-norm; FP32. score = logits[:,1] (bona fide).'
12
+ code: https://github.com/issflab/spoof_SUPERB/tree/8b457584e4120ba525f720e714d0a4b2ed400d68
13
+ checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/SpoofSUPERB-UniSpeechSAT/blob/1fd59fc73ab40845d54878b7667633f0b43b7b37/unispeech_sat_large.pth
14
+ params_millions: 315.698075
15
+ paper:
16
+ arxiv_id: '2603.01482'
17
+ url: https://arxiv.org/abs/2603.01482
18
+ bibtex: "@article{ali2026superbstyle,\n title={A SUPERB-Style Benchmark of Self-Supervised Speech\
19
+ \ Models for Audio Deepfake Detection},\n author={Ali, Hashim and Adupa, Nithin Sai and Subramani,\
20
+ \ Surya and Malik, Hafiz},\n journal={arXiv preprint arXiv:2603.01482},\n year={2026}\n}\n"
21
+ dataset:
22
+ id: SpeechAntiSpoofingBenchmarks/CFAD
23
+ revision: 53d7855c1c378524f7b7b1030bcb6b2caa327fe6
24
+ split: test
25
+ scores:
26
+ eer_percent: 16.443640173341585
27
+ n_trials: 62999
28
+ n_skipped: 0
29
+ artifact:
30
+ scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/SpoofSUPERB-UniSpeechSAT/resolve/865b030015598ac61bfd5a6f43525778932d1a19/.eval_results/SpeechAntiSpoofingBenchmarks/CFAD/scores.txt
31
+ scores_sha256: e3221887050e1005f4655c1b0e728bef529daaf754f05e8a45c5916e53e4a532
32
+ bench_version: speech-spoof-bench==0.4.1
33
+ reproduction: {}
34
+ submitter:
35
+ hf_username: korallll
36
+ contact: k.n.borodin@mtuci.ru
37
+ submitted_at: '2026-11-07'
38
+ notes: 'SpoofSUPERB-UniSpeechSAT = HF issf/spoof-superb-models @ 5245a543c8e5fa44eb87e5f353a3cb9f91f703b6
39
+ unispeech_sat_large.pth (downstream-only head: ssl_model.featurizer.weights[25], projector, post_net.linear;
40
+ sha256 198fbfb86fc016783c34d966a52a29b1f641eefae306f8b992d66a6ba0c2efce, matches the repo''s SHA256SUMS;
41
+ Apache-2.0) on the frozen upstream HF s3prl/converted_ckpts @ 8cad0b370e7e35f8d56951d95d2be036ea85510c
42
+ unispeech_sat_large.pt (sha256 5b2139cc20b0a649cc597cfa7f3872a6890a5ed23cd0163761da7c0ea408554e), i.e.
43
+ exactly what s3prl 0.4.18''s hub entry `unispeech_sat_large` downloads. Model = github.com/issflab/spoof_SUPERB
44
+ @ 8b457584 spoof_superb/models/linear_head.py (UtteranceLevel, MeanPooling, FrameLevel) + scoring/backends.py
45
+ (pad, logits[:,1]); the upstream is built with s3prl''s S3PRLUpstream/Featurizer from the local checkpoint
46
+ (no download); head loaded per backends.py (all 5 trained tensors present, nothing unexpected; upstream
47
+ loaded strict by s3prl). Verified bit-identical (max |d| = 0) to the authors'' code on In-the-Wild clips,
48
+ and within <0.01 logit of the authors'' published per-utterance In-the-Wild scores (issf/spoof-superb-scores)
49
+ when fed the original float wavs; the Arena''s PCM_16 FLAC copy of In-the-Wild moves logits slightly
50
+ more. Trained on ASVspoof 2019 LA train only (the Arena''s ASVspoof2019_LA is the eval split) -> no
51
+ trained_on entries. Self-reported EER% (paper main table): 19LA 1.35 / 21LA 8.82 / 21DF 7.59 / ASVspoof5
52
+ 15.00 / In-the-Wild 10.84.'