Add WST-Graph-Q82 submission for DECRO (EER=27.1274)

submissions/wst-graph-q82.yaml

@@ -0,0 +1,60 @@
+schema_version: 4
+system:
+  name: WST-Graph-Q82
+  slug: wst-graph-q82
+  description: 'WST-Graph (Ng et al., 2026), configuration Q82: a fixed, parameter-free Kymatio wavelet
+    scattering transform (J=8, Q=(8, 2), first + second order, oversampling 1) with log-modulation path
+    normalisation, arranged as a modulation-carrier grid (OrderGrid), length-aware attentive pooling to
+    64 relative-time bins (ALAP), a channel adapter + depthwise-separable latent blocks and an AASIST
+    graph back-end (GraphAASIST). Trained on ASVspoof 2019 LA train (focal loss, 12 epochs), training
+    seed 9 (the paper names no headline seed; seed 9 is the model card''s example). Authors'' traced TorchScript
+    program, run as released. Input as the model card: 16 kHz, no normalisation, first 64000 samples (4
+    s), right zero-pad, input_length = retained sample count. score = logit[bona fide] - logit[spoof]
+    (logits[:,0]-logits[:,1]). FP32 (TF32 off).'
+  code: https://huggingface.co/kwokho1/wst-graph/tree/080000c1685451c7759f43d2ae7d0f08477cfdf8
+  checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/WST-Graph-Q82/blob/b3a564c66f52b666330cb13a7864af912de30dd4/wst_graph_q82_seed9_batch32_cuda.ts
+  params_millions: 0.120364
+  paper:
+    arxiv_id: '2609.29372'
+    url: https://arxiv.org/abs/2609.29372
+    bibtex: "@article{ng2026wstgraph,\n  title   = {WST-Graph: Topology-Preserving Wavelet Scattering\
+      \ Front-End for Speech Deepfake Detection},\n  author  = {Ng, Kwok-Ho and Song, Tingting and Feng,\
+      \ Bingwen and Xia, Zhihua},\n  journal = {arXiv preprint arXiv:2609.29372},\n  year    = {2026}\n\
+      }\n"
+dataset:
+  id: SpeechAntiSpoofingBenchmarks/DECRO
+  revision: e54c29df5b81bac75e5cc294d1ab0d4db72a5c71
+  split: test
+scores:
+  eer_percent: 27.127402505669355
+  n_trials: 37314
+  n_skipped: 0
+artifact:
+  scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/WST-Graph-Q82/resolve/d281b3b1eb0ebe7548016ed7372d08e923d5999f/.eval_results/SpeechAntiSpoofingBenchmarks/DECRO/scores.txt
+  scores_sha256: 4e7305314dd476611c60bc244fff705e8dd37b3183ba1a090d8e1a79a1ad5218
+  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: 'WST-Graph-Q82 = HF kwokho1/wst-graph @ 080000c1685451c7759f43d2ae7d0f08477cfdf8, q82-seed9/wst_graph_q82_seed9_batch32_cuda.ts
+  (sha256 79edfcd396d14984c54ecbe08eae48e80605764291f004d2aa0cec0cf533f0f1; MIT, re-hosted). No Python
+  source is released (GitHub saki-ciallo/WST-Graph promised); the traced TorchScript program is the authors''
+  whole waveform-to-logits path. It was reviewed before running (code/ dir and model.code/graph of every
+  submodule): only aten::/prim:: ops, only the model''s own classes, pickles hold only tensors; the wrapper
+  re-asserts the aten/prim-only op set on load. The _cuda program has cuda:0 baked in (CUDA only). Despite
+  the "batch32" name it is batch-dynamic (verified B=1..48); B=1 takes a different cuFFT/GEMM path (|d
+  score| up to 0.0021 vs B>=2, which agree to ~4e-5), so the wrapper pads a lone clip with one zero clip
+  (input_length 1) and drops it. Seed choice: the paper (Table 6) reports single-seed results without
+  naming the seed and Tables 1-5 report three-seed averages, so seed 9 (model card example) is used for
+  both Q81 and Q82; not chosen on Arena data. Wrapper = model-card quick start exactly (same-batch |d|
+  = 0). Empty clips (the card raises) get a zero waveform with input_length 1. Trained on ASVspoof 2019
+  LA train (dev for selection); the Arena 19LA set is eval -> no trained_on entries. Self-reported EER%:
+  19LA 2.92 / ITW 44.46 / 21LA 8.53 / 21DF 18.15 / ASV5 35.55 / DFADD 21.56 / LibriSeVoc 33.58 / SONAR
+  39.53 (Table 6, single seed, Speech DF Arena protocol). ASVspoof2019_LA 2000-clip eval subset with this
+  wrapper: 3.033%. J-SPAW_LA scored with this wrapper (EER 43.644%).'