Add SpoofSUPERB-UniSpeechSAT submission for DECRO (EER=18.4446)

submissions/spoofsuperb-unispeechsat.yaml

@@ -0,0 +1,56 @@
+schema_version: 4
+system:
+  name: SpoofSUPERB-UniSpeechSAT
+  slug: spoofsuperb-unispeechsat
+  description: 'Spoof-SUPERB linear-head detector (Ali et al., 2026) over a frozen UniSpeech-SAT Large
+    (24 layers, 1024-d; s3prl `unispeech_sat_large`): the s3prl Featurizer softmax-weights all 25 hidden
+    states (24 transformer-layer inputs + the encoder output; no per-layer norm), then Linear(1024,256)
+    -> mean pooling over frames -> Linear(256,2). Only the layer weights, projector and classifier are
+    trained (ASVspoof 2019 LA train, 50 epochs, SWA); the authors'' released head with the s3prl converted
+    upstream. Input as the authors'' scoring code: 16 kHz, first 64600 samples (shorter clips tiled),
+    the upstream''s own per-utterance waveform layer-norm; FP32. score = logits[:,1] (bona fide).'
+  code: https://github.com/issflab/spoof_SUPERB/tree/8b457584e4120ba525f720e714d0a4b2ed400d68
+  checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/SpoofSUPERB-UniSpeechSAT/blob/1fd59fc73ab40845d54878b7667633f0b43b7b37/unispeech_sat_large.pth
+  params_millions: 315.698075
+  paper:
+    arxiv_id: '2603.01482'
+    url: https://arxiv.org/abs/2603.01482
+    bibtex: "@article{ali2026superbstyle,\n  title={A SUPERB-Style Benchmark of Self-Supervised Speech\
+      \ Models for Audio Deepfake Detection},\n  author={Ali, Hashim and Adupa, Nithin Sai and Subramani,\
+      \ Surya and Malik, Hafiz},\n  journal={arXiv preprint arXiv:2603.01482},\n  year={2026}\n}\n"
+dataset:
+  id: SpeechAntiSpoofingBenchmarks/DECRO
+  revision: e54c29df5b81bac75e5cc294d1ab0d4db72a5c71
+  split: test
+scores:
+  eer_percent: 18.444551128180507
+  n_trials: 37314
+  n_skipped: 0
+artifact:
+  scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/SpoofSUPERB-UniSpeechSAT/resolve/d151c5d7a96cd9279a01752e74bec94b039fc4d9/.eval_results/SpeechAntiSpoofingBenchmarks/DECRO/scores.txt
+  scores_sha256: 150a0598aece141117ee83ceba5ce400458820701ede086acbf5ae7de48fef13
+  bench_version: speech-spoof-bench==0.4.1
+reproduction:
+  reproduced_by: SpeechAntiSpoofingBenchmarks
+  reproduced_at: '2026-11-07'
+  reproduced_bench_version: speech-spoof-bench==0.4.1
+  match: inference
+submitter:
+  hf_username: korallll
+  contact: k.n.borodin@mtuci.ru
+submitted_at: '2026-11-07'
+notes: 'SpoofSUPERB-UniSpeechSAT = HF issf/spoof-superb-models @ 5245a543c8e5fa44eb87e5f353a3cb9f91f703b6
+  unispeech_sat_large.pth (downstream-only head: ssl_model.featurizer.weights[25], projector, post_net.linear;
+  sha256 198fbfb86fc016783c34d966a52a29b1f641eefae306f8b992d66a6ba0c2efce, matches the repo''s SHA256SUMS;
+  Apache-2.0) on the frozen upstream HF s3prl/converted_ckpts @ 8cad0b370e7e35f8d56951d95d2be036ea85510c
+  unispeech_sat_large.pt (sha256 5b2139cc20b0a649cc597cfa7f3872a6890a5ed23cd0163761da7c0ea408554e), i.e.
+  exactly what s3prl 0.4.18''s hub entry `unispeech_sat_large` downloads. Model = github.com/issflab/spoof_SUPERB
+  @ 8b457584 spoof_superb/models/linear_head.py (UtteranceLevel, MeanPooling, FrameLevel) + scoring/backends.py
+  (pad, logits[:,1]); the upstream is built with s3prl''s S3PRLUpstream/Featurizer from the local checkpoint
+  (no download); head loaded per backends.py (all 5 trained tensors present, nothing unexpected; upstream
+  loaded strict by s3prl). Verified bit-identical (max |d| = 0) to the authors'' code on In-the-Wild clips,
+  and within <0.01 logit of the authors'' published per-utterance In-the-Wild scores (issf/spoof-superb-scores)
+  when fed the original float wavs; the Arena''s PCM_16 FLAC copy of In-the-Wild moves logits slightly
+  more. Trained on ASVspoof 2019 LA train only (the Arena''s ASVspoof2019_LA is the eval split) -> no
+  trained_on entries. Self-reported EER% (paper main table): 19LA 1.35 / 21LA 8.82 / 21DF 7.59 / ASVspoof5
+  15.00 / In-the-Wild 10.84.'