Add Vocoded-VoxCeleb2-SSL-P3 submission for LRLspoof (1-SRR=31.7858)

submissions/vocoded-voxceleb2-ssl-p3.yaml

@@ -0,0 +1,65 @@
+schema_version: 4
+system:
+  name: Vocoded-VoxCeleb2-SSL-P3
+  slug: vocoded-voxceleb2-ssl-p3
+  description: 'NII vocoded-VoxCeleb2 SSL countermeasure, model ID P3 of Wang & Yamagishi (ICASSP 2024):
+    a single wav2vec 2.0 student front end initialised from XLSR-53 and distilled from the feature difference
+    of two teacher SSLs (v.vox minus XLSR-53), where v.vox is XLSR-53 continually pre-trained (self-supervised,
+    fairseq) on >9,000 h of VoxCeleb2 copy-synthesised by multiple neural vocoders; the paper''s proposed
+    single-SSL model. Back end: Linear 1024->128, three Linear+LeakyReLU+MC-dropout (p=0.5) blocks, mean
+    pooling over frames and Linear 128->2. Whole model fine-tuned on voc.LA (vocoded ASVspoof 2019 LA
+    train bona fide, ~10 h) with RawBoost. Score = logit of the mean bona-fide softmax over 144 Monte-Carlo-dropout
+    samples (the authors'' scoring), with fixed per-clip dropout masks. Authors'' released seed-01 checkpoint
+    (BSD-3). FP32, deterministic first-64600-sample window (shorter clips repeat-tiled) -- NOT the paper''s
+    full-utterance protocol.'
+  code: https://github.com/nii-yamagishilab/project-NN-Pytorch-scripts/tree/b529705b65588d62db17b56f98636ee734ff725e/project/10-asvspoof-vocoded-trn-ssl
+  checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/Vocoded-VoxCeleb2-SSL-P3/blob/e4405c9428c0c18bd7940bd0cb841d427280e3b6/model.safetensors
+  params_millions: 317.5716
+  paper:
+    arxiv_id: '2309.06014'
+    url: https://arxiv.org/abs/2309.06014
+    bibtex: "@inproceedings{wang2024vocoded,\n  title     = {Can Large-Scale Vocoded Spoofed Data Improve\
+      \ Speech Spoofing Countermeasure with a Self-Supervised Front End?},\n  author    = {Wang, Xin and\
+      \ Yamagishi, Junichi},\n  booktitle = {ICASSP 2024 - 2024 IEEE International Conference on Acoustics,\
+      \ Speech and Signal Processing (ICASSP)},\n  year      = {2024},\n  eprint    = {2309.06014},\n\
+      \  archivePrefix = {arXiv}\n}\n"
+dataset:
+  id: lab260/LRLspoof
+  revision: ac76f0e5472fdb68a6726e1fa94a148316e76916
+  split: test
+scores:
+  srr_complement: 31.785757827398132
+  n_trials: 1304455
+  n_skipped: 286
+calibration:
+  source_dataset: SpeechAntiSpoofingBenchmarks/DeepVoice
+  threshold: -2.859885917197452
+artifact:
+  scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/Vocoded-VoxCeleb2-SSL-P3/resolve/26b4aaf70bfaec476dc5bb7453abe50c628eff81/.eval_results/lab260/LRLspoof/scores.txt
+  scores_sha256: 1608374559ae7ac09f1eb2ad94c4116c9cbefb61da03d1a3ee96f3d13edc1230
+  bench_version: speech-spoof-bench==0.4.1
+reproduction:
+  reproduced_by: SpeechAntiSpoofingBenchmarks
+  reproduced_at: '2026-11-04'
+  reproduced_bench_version: speech-spoof-bench==0.4.1
+  match: inference
+submitter:
+  hf_username: korallll
+  contact: k.n.borodin@mtuci.ru
+submitted_at: '2026-11-04'
+notes: 'Code nii-yamagishilab/project-NN-Pytorch-scripts @ b529705 (BSD-3-Clause), project/10-asvspoof-vocoded-trn-ssl/model-ID-P3/model.py;
+  weights Zenodo record 8337778 project10-cm-ssl-model-ID-P3.tar, trained-for-paper/01/trained_network.pt
+  (a plain state_dict), re-saved unchanged as model.safetensors; strict load. P3: single SSL student front
+  end (init. XLSR-53), distilled (L1, weight 100) from the frozen P2-style teacher difference during fine-tuning.
+  The s3prl front end is rebuilt as a fairseq Wav2Vec2Model from the XLSR-53 architecture config (no hub
+  download), with s3prl''s per-utterance waveform layer_norm and last-layer (after final encoder LayerNorm)
+  features. Score = logit(mean over 144 MC-dropout samples of softmax[:, 1]) (bona fide = class 1; source
+  compute_llr_eps_ale, incl. its 1e-7 floors); dropout only in the back end, as in the source; the masks
+  are drawn per clip from a CUDA generator seeded by blake2b(seed 1, clip window), so scores are deterministic
+  and batch-independent (batched == single). Deviation (maintainer decision): fixed first 64600 samples,
+  repeat-tiled, instead of the authors'' full-utterance evaluation; consequently the source''s "< 3000
+  samples -> score 0" rule never applies. Full-utterance debug run of this port reproduces the authors''
+  released In-the-Wild scores (see implementation-notes.md). Released-log EERs (full utterance, seed 01):
+  19LA 1.73 / 21LA 14.88 / 21DF 6.02 / ITW 5.86. Training: SSL continual pre-training on vocoded VoxCeleb2
+  (unlabelled); CM fine-tuning on voc.LA (vocoded ASVspoof 2019 LA train bona fide), ASVspoof 2019 LA
+  dev for early stopping; no Arena test set used.'