Add Thaki-VoiceSpoof-W2V2-Large submission for LRLspoof (1-SRR=23.9639)

#43
submissions/thaki-voicespoof-w2v2-large.yaml ADDED
@@ -0,0 +1,62 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ schema_version: 4
2
+ system:
3
+ name: Thaki-VoiceSpoof-W2V2-Large
4
+ slug: thaki-voicespoof-w2v2-large
5
+ description: 'Thaki-VoiceSpoof-W2V2-Large (ThakiCloud, 2026, Research Preview): facebook/wav2vec2-large
6
+ (LibriSpeech 960 h self-supervised) with all 24 transformer layers and the feature projection fine-tuned
7
+ on ASVspoof 5 train (Track 1; CNN feature encoder frozen), plus softmax-weighted sum of the 25 hidden
8
+ states, Linear(1024,256) + GELU, attentive statistics pooling and a 2-way linear head. Per-utterance
9
+ zero-mean / unit-variance waveform normalisation inside the model. Authors'' released checkpoint (selected
10
+ epoch 5). Input exactly as the authors'' preprocess(): 16 kHz, first 64600 samples, shorter clips
11
+ tiled. Computed as the authors do: bf16 autocast (fp32 parameters), SDPA attention (cuDNN backend
12
+ where available), fixed 128-row batches (zero-padded, padded rows discarded). score = logit[bonafide]
13
+ - logit[spoof] (bf16, cast to float).'
14
+ code: https://huggingface.co/ThakiCloud/Thaki-VoiceSpoof-W2V2-Large/blob/093c7a1a8522670611c540f4912e86b356412821/voicespoof_inference.py
15
+ checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/Thaki-VoiceSpoof-W2V2-Large/blob/725e368ef4023eea5a226c3f08089bce8c469b9d/model.safetensors
16
+ params_millions: 315.757339
17
+ dataset:
18
+ id: lab260/LRLspoof
19
+ revision: ac76f0e5472fdb68a6726e1fa94a148316e76916
20
+ split: test
21
+ scores:
22
+ srr_complement: 23.96391878659898
23
+ n_trials: 1304455
24
+ n_skipped: 286
25
+ calibration:
26
+ source_dataset: SpeechAntiSpoofingBenchmarks/DeepVoice
27
+ threshold: 19.32301889278385
28
+ artifact:
29
+ scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/Thaki-VoiceSpoof-W2V2-Large/resolve/40def5d5161ddc85afc1f467c3da0e233bb4abc4/.eval_results/lab260/LRLspoof/scores.txt
30
+ scores_sha256: 0d510809ac910eb00c8322f8cf0e804ab715392a2270d07cd10f5e9886fdc047
31
+ bench_version: speech-spoof-bench==0.4.1
32
+ reproduction:
33
+ reproduced_by: SpeechAntiSpoofingBenchmarks
34
+ reproduced_at: '2026-11-16'
35
+ reproduced_bench_version: speech-spoof-bench==0.4.1
36
+ match: inference
37
+ submitter:
38
+ hf_username: korallll
39
+ contact: k.n.borodin@mtuci.ru
40
+ submitted_at: '2026-11-16'
41
+ notes: 'Thaki-VoiceSpoof-W2V2-Large = HF ThakiCloud/Thaki-VoiceSpoof-W2V2-Large @ 093c7a1a8522670611c540f4912e86b356412821,
42
+ model.safetensors (sha256 679b11aa005a52c5d649529dca673b6aff5f3fef751d994e1d980fddd66ca09b; 412 fp32
43
+ tensors, strict load, all used) + config.json (vendored as thaki_config.json; backbone built from it,
44
+ no Hub download). Apache-2.0 -> weights re-hosted. ASP + VoiceSpoofModel vendored verbatim from the
45
+ repo''s voicespoof_inference.py (read, never imported from the Hub cache) WITHOUT its runtime gate,
46
+ which only allows NVIDIA B200 + torch 2.9.0+cu130 + transformers 4.57.6. Our runtime: RTX 4070 Ti SUPER
47
+ / RTX 5060 Ti, torch 2.7.0+cu128, cuDNN 9.7, transformers 5.3.0, with the authors'' numerics kept: bf16
48
+ autocast, fp32 params, matmul TF32 off / cuDNN TF32 on, fixed 128-row batches. SDPA backend: cuDNN attention
49
+ (as the authors) on the RTX 4070 Ti SUPER; the RTX 5060 Ti (sm_120) has no cuDNN SDPA plan in cuDNN
50
+ 9.7, so there the default dispatcher (flash attention) runs. Verified against the authors'' VI.load
51
+ / VI.preprocess / VI.score path (gate bypassed in the test harness only) on 66 J-SPAW_LA clips: max
52
+ |d| = 0 on both cards (5060: their forced cuDNN kernel swapped for the same default kernel). bf16 scores
53
+ are coarse (1 bf16 step = 0.125 near the authors'' threshold 19.25; 430 distinct values on J-SPAW_LA)
54
+ and move by bf16 steps with the attention kernel / GPU (<= 0.25 on 256 clips; cuDNN vs flash max 2.7
55
+ on 1 of 2397 J-SPAW_LA clips) and, without the 128-row padding, with batch shape; the padding makes
56
+ scores independent of request splitting. Training: ASVspoof 5 train only (Arena ASVspoof5 is the eval
57
+ partition) -> no trained_on entries. No model paper (the card cites only the ASVspoof 5 dataset paper)
58
+ -> unpublished tier. Self-reported ASVspoof 5 eval EER 8.06% / minDCF 0.1995 (Dev EER 4.07%). J-SPAW_LA
59
+ EER 3.629% (PyTorch: RTX 5060 Ti 3.6287 / RTX 4070 Ti SUPER 3.6288, Spearman 0.99998 between cards).
60
+ Scored with PyTorch, not TensorRT: a BF16 engine passes the Arena TRT rule (J-SPAW_LA EER 3.703 / 3.753,
61
+ Spearman 0.9998) but moves single scores by up to 7.4 points away from the authors'' bf16 autocast numerics
62
+ for a 7-37 % speed-up.'