Add MMS-300M-AntiDeepfake-NDA submission for ADD22_eval_31 (EER=3.5559) (#14)

- Add MMS-300M-AntiDeepfake-NDA submission for ADD22_eval_31 (EER=3.5559) (9567aeda2c090e7e40251906fbcbcff1e6c21429)
- Fill reproduction block (inference, maintainer re-ran the model) (45eadfb3de323bd2a24813e7d79b91844e777a53)

submissions/mms-300m-antideepfake-nda.yaml

@@ -0,0 +1,45 @@
+schema_version: 4
+system:
+  name: MMS-300M-AntiDeepfake-NDA
+  slug: mms-300m-antideepfake-nda
+  description: 'AntiDeepfake detector (MMS-300M backbone): a self-supervised speech foundation model post-trained
+    for deepfake/spoof detection, with an AdaptiveAvgPool1d + single fully-connected layer head over the
+    SSL features. Released by nii-yamagishilab (arXiv:2506.21090). Trained **without** data augmentation
+    (NDA variant). Checkpoint copied bit-exactly from nii-yamagishilab/mms-300m-anti-deepfake-nda. Evaluated
+    here with a deterministic first-4 s window and batching (NOT the paper''s whole-utterance protocol).'
+  code: https://github.com/nii-yamagishilab/AntiDeepfake
+  checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/MMS-300M-AntiDeepfake-NDA
+  params_millions: 315.0
+  paper:
+    arxiv_id: '2506.21090'
+    url: https://arxiv.org/abs/2506.21090
+    bibtex: "@article{ge2025posttraining,\n  title={Post-training for Deepfake Speech Detection},\n  author={Ge,\
+      \ Wanying and Wang, Xin and Kawahara, Tomoki and Yamamoto, Yuto and Delgado, Hector and Todisco,\
+      \ Massimiliano and Evans, Nicholas and Yamagishi, Junichi},\n  journal={arXiv preprint arXiv:2506.21090},\n\
+      \  year={2025}\n}\n"
+dataset:
+  id: SpeechAntiSpoofingBenchmarks/ADD22_eval_31
+  revision: a55a9f8f6bbab3e41dd8f354549aa03c949be2e9
+  split: test
+scores:
+  eer_percent: 3.555901434482965
+  n_trials: 112861
+  n_skipped: 0
+artifact:
+  scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/MMS-300M-AntiDeepfake-NDA/resolve/82be6a6d21b59828034420eac1c4688a74af2957/.eval_results/SpeechAntiSpoofingBenchmarks/ADD22_eval_31/scores.txt
+  scores_sha256: 70a2bc6397a347c0bb54dc1fffb9e5c65ab5c22693d5e055d7cb3f306f4d9e65
+  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: inference
+submitter:
+  hf_username: korallll
+  contact: k.n.borodin@mtuci.ru
+submitted_at: '2026-10-29'
+notes: 'score = output logit for class 1 (real / bona fide); higher = more bona fide. Backbone built from
+  the source fairseq config (mms_300m); the released model.safetensors overwrites every weight (220-key
+  exact match family-wide). Input: fixed first 64000 samples (4.0 s @ 16 kHz), tile-repeated if shorter,
+  per-utterance layer_norm; no resampling. Checkpoint license: CC BY-NC-SA 4.0. STRICTLY PROHIBITED on
+  the model''s in-domain corpora (see README); Arena runs ArAD + In-the-Wild only.'