Add NII-SLT26-XLSR-FC submission for PyAra (EER=30.8187)

submissions/nii-slt26-xlsr-fc.yaml

@@ -0,0 +1,61 @@
+schema_version: 4
+system:
+  name: NII-SLT26-XLSR-FC
+  slug: nii-slt26-xlsr-fc
+  description: 'XLS-R + FC from Ge, Wang and Yamagishi (IEEE SLT 2026): a fully fine-tuned wav2vec 2.0
+    XLS-R 300M encoder (transformers Wav2Vec2Model), mean pooling of the last hidden state over all frames,
+    and a Linear(1024, 2) head [real, fake]. Trained on CompSpoof V2 (ESDD2) train with the "any component
+    fake = fake" labels, the unmixed "original" class excluded (protocol AnyFakeNoOriginal), seed 2, best-validation-loss
+    checkpoint. Authors'' released checkpoint. Input as the authors'' collator: 16 kHz waveform, per-clip
+    zero-mean unit-variance normalisation (Wav2Vec2FeatureExtractor), applied after the Arena 4 s cap
+    (first 64600 samples, shorter clips repeat-padded; the authors score whole clips). score = logit[real]
+    - logit[fake] (the authors'' Score_real - Score_fake on softmax probabilities is a monotone function
+    of it). FP32 (the authors'' eval precision; TF32 off).'
+  code: https://github.com/nii-yamagishilab/AntiDeepfake-AudioLLM/tree/2d445b525bb52518448b353fe1fee51d5f815af9
+  checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/NII-SLT26-XLSR-FC/blob/60b2822b59f136d86644e9f251092c49265569d4/nii_slt26_xlsr_fc.safetensors
+  params_millions: 315.439746
+  paper:
+    arxiv_id: '2610.17617'
+    url: https://arxiv.org/abs/2610.17617
+    bibtex: "@inproceedings{ge2026what,\n  title     = {What Do Deepfake Audio Detectors Hear? From Self-Supervised\
+      \ Models to Audio LLMs},\n  author    = {Ge, Wanying and Wang, Xin and Yamagishi, Junichi},\n  booktitle\
+      \ = {IEEE Spoken Language Technology Workshop (SLT)},\n  year      = {2026}\n}\n"
+dataset:
+  id: SpeechAntiSpoofingBenchmarks/PyAra
+  revision: 40167badd6d8a289f02054761ee5411f49227b87
+  split: test
+scores:
+  eer_percent: 30.818674675299352
+  n_trials: 201778
+  n_skipped: 0
+artifact:
+  scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/NII-SLT26-XLSR-FC/resolve/840fe1942ec6a5e4aa5ebe944652313310e09415/.eval_results/SpeechAntiSpoofingBenchmarks/PyAra/scores.txt
+  scores_sha256: 5d7ba0f5779578491ccdf0add406c31e872ce79f1bb5b491c93f28274d728a9b
+  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: 'HF nii-yamagishilab/slt2026-checkpoints @ ae9d2bc72b670e49513f5bd125bf679095bf4b92, checkpoints.tar.gz
+  -> xlsr-fc.ckpt (sha256 0caa8335c5e5504d114f24415d20c8f0082a5b96c3b9c81dd3fe913d7db8a5bc; a torch training
+  checkpoint {step 79205, epoch 5, model_state_dict, optimizer_state_dict, config}; pickle globals inspected
+  with pickletools, loaded with weights_only=True). model_state_dict re-saved unchanged as nii_slt26_xlsr_fc.safetensors
+  (424 tensors, sha256 3dc2417582c6cac8287d7bdfe6fb5224de001dab19f46d38409e76dc2820c47a) + the needed
+  config values as JSON; re-hosted under CC BY-NC 4.0 (LICENSE-CHECKPOINT). Code BSD-3. The XLS-R encoder
+  is built from the facebook/wav2vec2-xls-r-300m config (HF @ 1a640f3) and every tensor comes from the
+  checkpoint (strict load, all 424 tensors used). Labels: real id 0, fake id 1 (checkpoint config). Verified
+  against the authors'' own build_encoder_fc_components + EncoderFCCollator + ClassificationHead on 18
+  J-SPAW_LA clips: max |d logit-diff| 2.4e-6 with the 4 s cap (their batch 1 and 8), 0.0 with the window
+  off. Deviation: Arena 4 s cap instead of whole clips (random 30 s crop if longer). Training data CompSpoof
+  V2 train: bona fide speech CommonVoice, LibriTTS, english-conversation-corpus; spoof speech ASVspoof
+  5 and MLAAD; environment AudioCaps, TAU UAS, TUT SED, UrbanSound, VGGSound, EnvSDD, VcapAV (arXiv 2606.10791
+  Table II; dataset gated, splits unstated) -> LibriTTS and ASVspoof 5 taken as whole corpora -> ASVspoof5,
+  SONAR, CD-ADD, LibriSeVoc marked trained_on (n/a*). Every training bona fide clip is speech mixed with
+  an environmental sound, so clean studio speech is out of domain (most Arena clips score as fake). Self-reported
+  CompSpoof V2 test EER% (3-seed mean): fake event 9.50 / fake speech 21.21 / both 13.96. J-SPAW_LA EER
+  36.756% (PyTorch); TensorRT FP32 on RTX 4070 Ti SUPER 36.875% (Spearman 0.99998 vs PyTorch).'