Add Pellav3 submission for PyAra (EER=0.1277) (#72)

- Add Pellav3 submission for PyAra (EER=0.1277) (5c9651631f54535ba2acd433d3d9cb6c048163c3)
- Fill reproduction block (inference, maintainer re-ran the model) (34eb5be946c9ffcd6b3a39411327e033ecb37a21)

submissions/pellav3.yaml

@@ -0,0 +1,55 @@
+schema_version: 4
+system:
+  name: Pellav3
+  slug: pellav3
+  description: 'Pellav3 (Sadanie, November 2026): an open-weights two-view audio deepfake detector (1.54
+    B parameters). Each 4 s window is read by XLS-R 1B on the standardised waveform (softmax mix of all
+    49 hidden states) and by w2v-BERT 2.0 on 80-bin kaldi log-mel fbank features (25-state mix); each
+    view is projected to 256-d and refined by two Res2 blocks, the frame sequences are concatenated per
+    frame (512-d), refined by two shared Res2 blocks and pooled by attentive statistics pooling into Linear(1024,1)
+    -> logit(fake). Authors'' released checkpoint; training data undisclosed. Input as the authors'' pellav3_infer.py
+    view(): 16 kHz, mean removal, 20 Hz high-pass (zero-phase 2nd-order Butterworth), ITU-R BS.1770 loudness
+    normalisation to -23 LUFS (pyloudnorm semantics), zero-pad to 4 s. Arena 4 s cap: one window, the
+    first 4 s (64000 samples) of the authors'' window grid (the authors average 4 s windows every 2 s
+    over longer clips; clips <= 4 s are scored identically). score = -logit(fake), higher = bona fide.
+    FP32 (TF32 off; the authors run bfloat16 autocast).'
+  code: https://huggingface.co/Sadanie/pellav3-audio-deepfake-detector/blob/47b69420803c01de65af264ecdc2c5f35f2184f3/pellav3_infer.py
+  checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/Pellav3/blob/9bb345f5ec6dedc34f26458d5a5d545d9bb1c3ab/pellav3_detector.safetensors
+  params_millions: 1544.093452
+dataset:
+  id: SpeechAntiSpoofingBenchmarks/PyAra
+  revision: 40167badd6d8a289f02054761ee5411f49227b87
+  split: test
+scores:
+  eer_percent: 0.12774689806069336
+  n_trials: 201778
+  n_skipped: 0
+artifact:
+  scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/Pellav3/resolve/db0857f2ab17e7da0d3e3573b095c00a0dc4fa2e/.eval_results/SpeechAntiSpoofingBenchmarks/PyAra/scores.txt
+  scores_sha256: 3abe8a2d5e07ff3124000d6642f25e2718cd3713c310a7a15ee82dc383ed9ee9
+  bench_version: speech-spoof-bench==0.4.1
+reproduction:
+  reproduced_by: SpeechAntiSpoofingBenchmarks
+  reproduced_at: '2026-11-11'
+  reproduced_bench_version: speech-spoof-bench==0.4.1
+  match: inference
+submitter:
+  hf_username: korallll
+  contact: k.n.borodin@mtuci.ru
+submitted_at: '2026-11-11'
+notes: 'Pellav3 = HF Sadanie/pellav3-audio-deepfake-detector @ 47b69420803c01de65af264ecdc2c5f35f2184f3,
+  pellav3_detector.pt (sha256 ac6ae020ac5de06beee4f283ad1ea1fb1ae88dd8da19f483a602a032966695ec; 1647 fp32
+  tensors, weights_only=True, strict) re-hosted unchanged as pellav3_detector.safetensors (sha256 6a721d66a0b9e1a1011193a66f245a4836dbfc036eb671cd1ff80d6e266d5a73).
+  MIT per the model card (no LICENSE file). Model classes and preprocessing vendored from pellav3_infer.py
+  (sha256 c38f9a43..., read, never imported from the Hub cache); backbones built from the facebook/wav2vec2-xls-r-1b
+  (@ 35eaea9a) and facebook/w2v-bert-2.0 (@ da985ba0) configs, as the script does, every weight from the
+  checkpoint. The script pins transformers 4.48.3; backbone hidden states are bit-identical under 4.49.0
+  and the 5.3.0 used here. pyloudnorm''s integrated loudness is re-implemented (not installed here). One
+  edit: XLS-R frames past each clip''s own w2v-BERT length are zeroed so a padded batch equals one-clip-at-a-time
+  scoring (no-op for a single clip). Verified against the authors'' Detector + decode/view on 20 J-SPAW_LA-derived
+  clips (FP32): first-window max |d| 7.6e-6, full multi-window path max |d| 0.0; the authors'' own bf16
+  path differs by <= 0.18 logit on the same window. Deviations: Arena 4 s cap (first window only instead
+  of the mean over 4 s windows every 2 s; identical for clips <= 4 s), FP32 instead of bf16 autocast.
+  Training data is undisclosed (no paper, card silent) -> no trained_on entries, but leakage cannot be
+  ruled out: J-SPAW_LA EER is 0.0% with saturated, perfectly separated scores. No paper -> unpublished
+  tier.'