Add ProSDD-ASV5 submission for LRLspoof (1-SRR=35.7487)

#30
by korallll - opened
Files changed (1) hide show
  1. submissions/prosdd-asv5.yaml +62 -0
submissions/prosdd-asv5.yaml ADDED
@@ -0,0 +1,62 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ schema_version: 4
2
+ system:
3
+ name: ProSDD-ASV5
4
+ slug: prosdd-asv5
5
+ description: 'ProSDD (Mahapatra et al., 2026): a fully fine-tuned wav2vec 2.0 XLS-R 300M encoder, first
6
+ trained (Stage I, bona fide LibriSpeech train-clean-100 + dev only) to predict masked frame-level
7
+ speaker + prosody targets, then (Stage II) jointly fine-tuned with that objective and a binary real/fake
8
+ head on ASVspoof 2024 (ASVspoof 5) train/dev (RawBoost series 3). Inference uses only the clean path:
9
+ XLS-R CNN front-end + feature projection, latent sequence fixed at 200 frames (zero-padded), 24-layer
10
+ transformer encoder, mean pool, Linear(1024,512)-ReLU-Linear(512,2). Input as the authors'' data_utils_eval.py:
11
+ 16 kHz, no amplitude normalisation, fixed 4 s (64000 samples) centre crop, shorter clips centre zero-padded.
12
+ FP32; score = logits[:,1] (bona fide class). Authors'' released checkpoint; no upstream license, so
13
+ the weights are linked from the authors'' Google Drive, not re-hosted.'
14
+ code: https://github.com/ProSDD/codes/tree/bdb80440cd946368e77f4b51f6ddf26c7443dc33
15
+ checkpoint: https://drive.google.com/file/d/1AyD4f0RVVyDncqYbhi8QaP6-GFZ_ZGZl/view
16
+ params_millions: 315.963522
17
+ paper:
18
+ arxiv_id: '2604.13229'
19
+ url: https://arxiv.org/abs/2604.13229
20
+ bibtex: "@article{mahapatra2026prosdd,\n title = {ProSDD: Learning Prosodic Representations for\
21
+ \ Speech Deepfake Detection against Expressive and Emotional Attacks},\n author = {Mahapatra,\
22
+ \ Aurosweta and Ulgen, Ismail Rasim and Lee, Kong Aik and Andrews, Nicholas and Sisman, Berrak},\n\
23
+ \ journal = {arXiv preprint arXiv:2604.13229},\n year = {2026}\n}\n"
24
+ dataset:
25
+ id: lab260/LRLspoof
26
+ revision: ac76f0e5472fdb68a6726e1fa94a148316e76916
27
+ split: test
28
+ scores:
29
+ srr_complement: 35.748741152412
30
+ n_trials: 1304455
31
+ n_skipped: 286
32
+ calibration:
33
+ source_dataset: SpeechAntiSpoofingBenchmarks/DeepVoice
34
+ threshold: 7.902336933559322
35
+ artifact:
36
+ scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/ProSDD-ASV5/resolve/02167ad8b1717a275d241a5d3129872ebcc30049/.eval_results/lab260/LRLspoof/scores.txt
37
+ scores_sha256: e2efb38708a6d052ac78b9657445a5f6598744cf26261cdab98d34ac45b45432
38
+ bench_version: speech-spoof-bench==0.4.1
39
+ reproduction:
40
+ reproduced_by: SpeechAntiSpoofingBenchmarks
41
+ reproduced_at: '2026-11-09'
42
+ reproduced_bench_version: speech-spoof-bench==0.4.1
43
+ match: inference
44
+ submitter:
45
+ hf_username: korallll
46
+ contact: k.n.borodin@mtuci.ru
47
+ submitted_at: '2026-11-09'
48
+ notes: 'ProSDD-ASV5 = github.com/ProSDD/codes @ bdb80440cd946368e77f4b51f6ddf26c7443dc33 (model_stage2realfake.py
49
+ ProSDDStage2, main_eval.py inference_forward, data_utils_eval.py pad), checkpoint ProSDD_Trained2024
50
+ from the authors'' Google Drive folder 1h250-Um5qWo-rpeOE6K_Gdeygy46k3xg (file id 1AyD4f0RVVyDncqYbhi8QaP6-GFZ_ZGZl,
51
+ sha256 c044a402d190f7d4bf75165e56890c21c0d80ffe4164ad741f4ea7d48356c437; a plain PyTorch state_dict,
52
+ loaded with weights_only=True). The repo has no LICENSE, so the weights are linked, not re-hosted. The
53
+ XLS-R 300M backbone is built from the facebook/wav2vec2-xls-r-300m config (HF @ 1a640f3) and every tensor
54
+ is overwritten by the checkpoint (strict load; the 5 training-only tensors mask_embed, pros_ln.*, final_proj.*
55
+ (Stage-I 1024->448 speaker+prosody head) are dropped via an asserted allow-list; the source''s own strict=False
56
+ load reports 0 missing / 0 unexpected). classifier_pool = mean (the checkpoint has no attention-pool
57
+ tensors). Preprocessing exactly as data_utils_eval.py / main_eval.py: 64000-sample centre crop / centred
58
+ zero-pad, no normalisation, latent frames 199 -> zero-padded to 200, encoder without attention mask,
59
+ mean over all 200 frames. Score = logits[:,1]; verified against the authors'' modules (max |d| 1e-5).
60
+ Training: Stage I LibriSpeech train-clean-100 + dev (bona fide), Stage II ASVspoof 2024 (ASVspoof 5)
61
+ train/dev with RawBoost 3. The Arena ASVspoof sets are eval partitions, so no trained_on entries. Self-reported
62
+ EER% (paper): 19LA 19.04 / 21LA 18.08 / ASVspoof 2024 (ASVspoof5) 7.38 / EmoFake 25.06.'