Add Forensics-WavLM-OCSoftmax submission for DECRO (EER=2.7844)

submissions/forensics-wavlm-ocsoftmax.yaml

@@ -0,0 +1,51 @@
+schema_version: 4
+system:
+  name: Forensics-WavLM-OCSoftmax
+  slug: forensics-wavlm-ocsoftmax
+  description: 'Forensics 0.3B WavLM OC-Softmax speech deepfake detector (Eliya Segev, 2026): a fully
+    fine-tuned WavLM-Large front-end whose 24 transformer hidden states are pooled by an AASIST-style
+    spectro-temporal graph-attention pool (layer x time graphs, GAT dims 128/64), followed by a 3-block
+    residual MLP classifier (320-d) and a single logit (label 1 = real); trained with a combined CE +
+    OC-Softmax objective (run wavlm_aasist_27_03_oc_softmax_ce_from_16, last epoch). Authors'' 2nd_oc_softmax_ce_from16_last
+    checkpoint as released. Input as the authors'' inference.py: peak-normalised waveform, 5 s (80000
+    samples) centre crop, shorter clips repeat-padded. score = the logit (sigmoid = P(real)); the OC-Softmax
+    centre is not shipped. Unpublished model (HF model card only, no paper).'
+  code: https://huggingface.co/eliya/forensics_0.3B_wavlm_oc_softmax_deepfake_classifier/tree/2911b8f1f6112e7095c8ac1061e03b81baff95a2
+  checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/Forensics-WavLM-OCSoftmax/blob/c620def36d1cde2845dafda243bbd9ed8ca7e525/2nd_oc_softmax_ce_from16_last.safetensors
+  params_millions: 317.257863
+dataset:
+  id: SpeechAntiSpoofingBenchmarks/DECRO
+  revision: e54c29df5b81bac75e5cc294d1ab0d4db72a5c71
+  split: test
+scores:
+  eer_percent: 2.784445511281805
+  n_trials: 37314
+  n_skipped: 0
+artifact:
+  scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/Forensics-WavLM-OCSoftmax/resolve/452b25308d4d361d56c6a914401f9a9d5887c8fc/.eval_results/SpeechAntiSpoofingBenchmarks/DECRO/scores.txt
+  scores_sha256: ccd7bbf64f97118947e89ab9e63f6928f045c27a4071f929c416c50ef827f56c
+  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: 'Forensics-WavLM-OCSoftmax = HF eliya/forensics_0.3B_wavlm_oc_softmax_deepfake_classifier @ 2911b8f1f6112e7095c8ac1061e03b81baff95a2,
+  2nd_oc_softmax_ce_from16_last.safetensors (sha256 53ed62c9...f9a), CC-BY-NC-4.0, re-hosted unchanged.
+  Architecture = the repo''s model.py (byte-identical to Forensics-WavLM-Base''s); the WavLM-Large backbone
+  is built from the microsoft/wavlm-large config and every weight is overwritten by the checkpoint (strict
+  load, incl. the weight-normed pos_conv; the 4 training-only projection.* tensors, unused by model.py,
+  are dropped via an asserted allow-list). Preprocessing = inference.py: peak normalisation wav/(max|wav|+1e-8)
+  over the whole clip, then a centred 80000-sample (5 s) crop or repeat-padding; no resampling; FP32.
+  Score = the single output logit (higher = bona fide), as in the authors'' inference.py; verified bit-identical
+  to the authors'' model.py + inference.py on 20 In-the-Wild clips. The card says the "last" epoch (10
+  of the sub-run) was shipped rather than the metric-best one. Training data: the current model card says
+  5M+ samples from a mix of proprietary and public data; earlier card revisions listed SpeechFake, DFADD,
+  CodecFake, MD-CommonVoice, ASVspoof2019-LA (later ones add ASVspoof2021, CommonVoice, Aishell); splits
+  are not stated, so InTheWild, DFADD, ASVspoof2021_DF and ASVspoof2019_LA are marked trained_on (n/a*),
+  as for Forensics-XLSR-Wild. Self-reported EER: In-the-Wild 1.76%, ASVspoof2019-LA 0.68%. No paper ->
+  Unpublished tier.'