Add Thaki-VoiceSpoof-W2V2-Large submission for DECRO (EER=43.2668)

submissions/thaki-voicespoof-w2v2-large.yaml

@@ -0,0 +1,59 @@
+schema_version: 4
+system:
+  name: Thaki-VoiceSpoof-W2V2-Large
+  slug: thaki-voicespoof-w2v2-large
+  description: 'Thaki-VoiceSpoof-W2V2-Large (ThakiCloud, 2026, Research Preview): facebook/wav2vec2-large
+    (LibriSpeech 960 h self-supervised) with all 24 transformer layers and the feature projection fine-tuned
+    on ASVspoof 5 train (Track 1; CNN feature encoder frozen), plus softmax-weighted sum of the 25 hidden
+    states, Linear(1024,256) + GELU, attentive statistics pooling and a 2-way linear head. Per-utterance
+    zero-mean / unit-variance waveform normalisation inside the model. Authors'' released checkpoint (selected
+    epoch 5). Input exactly as the authors'' preprocess(): 16 kHz, first 64600 samples, shorter clips
+    tiled. Computed as the authors do: bf16 autocast (fp32 parameters), SDPA attention (cuDNN backend
+    where available), fixed 128-row batches (zero-padded, padded rows discarded). score = logit[bonafide]
+    - logit[spoof] (bf16, cast to float).'
+  code: https://huggingface.co/ThakiCloud/Thaki-VoiceSpoof-W2V2-Large/blob/093c7a1a8522670611c540f4912e86b356412821/voicespoof_inference.py
+  checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/Thaki-VoiceSpoof-W2V2-Large/blob/725e368ef4023eea5a226c3f08089bce8c469b9d/model.safetensors
+  params_millions: 315.757339
+dataset:
+  id: SpeechAntiSpoofingBenchmarks/DECRO
+  revision: e54c29df5b81bac75e5cc294d1ab0d4db72a5c71
+  split: test
+scores:
+  eer_percent: 43.26677152451822
+  n_trials: 37314
+  n_skipped: 0
+artifact:
+  scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/Thaki-VoiceSpoof-W2V2-Large/resolve/51d00f48dd7d453f676ef0300e8046cfe9dee363/.eval_results/SpeechAntiSpoofingBenchmarks/DECRO/scores.txt
+  scores_sha256: 9981ccf08ccf8721dae856ccc1f6eb9f1714b98c03ccf792fb7079c038b51f82
+  bench_version: speech-spoof-bench==0.4.1
+reproduction:
+  reproduced_by: SpeechAntiSpoofingBenchmarks
+  reproduced_at: '2026-11-10'
+  reproduced_bench_version: speech-spoof-bench==0.4.1
+  match: inference
+submitter:
+  hf_username: korallll
+  contact: k.n.borodin@mtuci.ru
+submitted_at: '2026-11-10'
+notes: 'Thaki-VoiceSpoof-W2V2-Large = HF ThakiCloud/Thaki-VoiceSpoof-W2V2-Large @ 093c7a1a8522670611c540f4912e86b356412821,
+  model.safetensors (sha256 679b11aa005a52c5d649529dca673b6aff5f3fef751d994e1d980fddd66ca09b; 412 fp32
+  tensors, strict load, all used) + config.json (vendored as thaki_config.json; backbone built from it,
+  no Hub download). Apache-2.0 -> weights re-hosted. ASP + VoiceSpoofModel vendored verbatim from the
+  repo''s voicespoof_inference.py (read, never imported from the Hub cache) WITHOUT its runtime gate,
+  which only allows NVIDIA B200 + torch 2.9.0+cu130 + transformers 4.57.6. Our runtime: RTX 4070 Ti SUPER
+  / RTX 5060 Ti, torch 2.7.0+cu128, cuDNN 9.7, transformers 5.3.0, with the authors'' numerics kept: bf16
+  autocast, fp32 params, matmul TF32 off / cuDNN TF32 on, fixed 128-row batches. SDPA backend: cuDNN attention
+  (as the authors) on the RTX 4070 Ti SUPER; the RTX 5060 Ti (sm_120) has no cuDNN SDPA plan in cuDNN
+  9.7, so there the default dispatcher (flash attention) runs. Verified against the authors'' VI.load
+  / VI.preprocess / VI.score path (gate bypassed in the test harness only) on 66 J-SPAW_LA clips: max
+  |d| = 0 on both cards (5060: their forced cuDNN kernel swapped for the same default kernel). bf16 scores
+  are coarse (1 bf16 step = 0.125 near the authors'' threshold 19.25; 430 distinct values on J-SPAW_LA)
+  and move by bf16 steps with the attention kernel / GPU (<= 0.25 on 256 clips; cuDNN vs flash max 2.7
+  on 1 of 2397 J-SPAW_LA clips) and, without the 128-row padding, with batch shape; the padding makes
+  scores independent of request splitting. Training: ASVspoof 5 train only (Arena ASVspoof5 is the eval
+  partition) -> no trained_on entries. No model paper (the card cites only the ASVspoof 5 dataset paper)
+  -> unpublished tier. Self-reported ASVspoof 5 eval EER 8.06% / minDCF 0.1995 (Dev EER 4.07%). J-SPAW_LA
+  EER 3.629% (PyTorch: RTX 5060 Ti 3.6287 / RTX 4070 Ti SUPER 3.6288, Spearman 0.99998 between cards).
+  Scored with PyTorch, not TensorRT: a BF16 engine passes the Arena TRT rule (J-SPAW_LA EER 3.703 / 3.753,
+  Spearman 0.9998) but moves single scores by up to 7.4 points away from the authors'' bf16 autocast numerics
+  for a 7-37 % speed-up.'