Add SONAR-Full submission for LRLspoof (1-SRR=21.4417)

submissions/sonar-full.yaml

@@ -0,0 +1,51 @@
+schema_version: 4
+system:
+  name: SONAR-Full
+  slug: sonar-full
+  description: 'SONAR (Spectral-Contrastive Audio Residuals): a dual-path detector with two wav2vec 2.0
+    XLS-R 300M encoders. One encodes the raw waveform (content / low-frequency path); the other encodes
+    a high-frequency residual from a learnable constrained SRM filter bank (30 filters, k=5). The two
+    feature streams are fused by bidirectional cross-attention, projected to 128-d, and classified by
+    an AASIST spectro-temporal graph-attention back-end; training adds a Jensen-Shannon alignment loss
+    between the two paths. Official SONAR-Full checkpoint (sonar_full_xlsr_aasist_eer6.pth), trained on
+    ASVspoof2019 LA, FP32, deterministic first-64600-sample window (zero-padded).'
+  code: https://github.com/idonithid/SONAR-Audio-DF-Detection
+  checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/SONAR-Full
+  params_millions: 641.5252
+  paper:
+    arxiv_id: '2511.21325'
+    url: https://proceedings.mlr.press/v306/hidekel26a.html
+    bibtex: "@inproceedings{hidekel2026sonar,\n  title     = {{SONAR}: Spectral-Contrastive Audio Residuals\
+      \ for Generalizable Deepfake Detection},\n  author    = {Hidekel, Ido Nitzan and Lifshitz, Gal and\
+      \ Cohen, Khen and Raviv, Dan},\n  booktitle = {Proceedings of the 43rd International Conference\
+      \ on Machine Learning (ICML)},\n  series    = {Proceedings of Machine Learning Research},\n  volume\
+      \    = {306},\n  pages     = {42998--43012},\n  year      = {2026}\n}\n"
+dataset:
+  id: lab260/LRLspoof
+  revision: ac76f0e5472fdb68a6726e1fa94a148316e76916
+  split: test
+scores:
+  srr_complement: 21.44169043333689
+  n_trials: 1304455
+  n_skipped: 458
+calibration:
+  source_dataset: SpeechAntiSpoofingBenchmarks/DeepVoice
+  threshold: 1.1537789609039548
+artifact:
+  scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/SONAR-Full/resolve/8a2e366eb18b15df7ba22d0441df104b1e8e54d2/.eval_results/lab260/LRLspoof/scores.txt
+  scores_sha256: a30074521a8f165e9b59d81c365aa6b249a5f999677aa5bbeed2331e31584d2b
+  bench_version: speech-spoof-bench==0.4.1
+reproduction:
+  reproduced_by: SpeechAntiSpoofingBenchmarks
+  reproduced_at: '2026-10-29'
+  reproduced_bench_version: speech-spoof-bench==0.4.1
+  match: scoring
+submitter:
+  hf_username: korallll
+  contact: k.n.borodin@mtuci.ru
+submitted_at: '2026-10-29'
+notes: 'SONAR-Full from idonithid/SONAR-Audio-DF-Detection (ICML 2026): dual XLS-R 300M encoders (raw
+  waveform + constrained-SRM high-frequency residual) fused by cross-attention, AASIST back-end. Official
+  sonar_full_xlsr_aasist_eer6.pth weights (CC-BY-NC-4.0), strict load. Deterministic first-64600-sample
+  window, zero-padded for shorter clips (source sonar/inference.py). score = output logit for class 1
+  (real); higher = more bona fide.'