Add NII-SLT26-XLSR-FC submission for LRLspoof (1-SRR=50.7161)

#36
by korallll - opened
Files changed (1) hide show
  1. submissions/nii-slt26-xlsr-fc.yaml +64 -0
submissions/nii-slt26-xlsr-fc.yaml ADDED
@@ -0,0 +1,64 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ schema_version: 4
2
+ system:
3
+ name: NII-SLT26-XLSR-FC
4
+ slug: nii-slt26-xlsr-fc
5
+ description: 'XLS-R + FC from Ge, Wang and Yamagishi (IEEE SLT 2026): a fully fine-tuned wav2vec 2.0
6
+ XLS-R 300M encoder (transformers Wav2Vec2Model), mean pooling of the last hidden state over all frames,
7
+ and a Linear(1024, 2) head [real, fake]. Trained on CompSpoof V2 (ESDD2) train with the "any component
8
+ fake = fake" labels, the unmixed "original" class excluded (protocol AnyFakeNoOriginal), seed 2, best-validation-loss
9
+ checkpoint. Authors'' released checkpoint. Input as the authors'' collator: 16 kHz waveform, per-clip
10
+ zero-mean unit-variance normalisation (Wav2Vec2FeatureExtractor), applied after the Arena 4 s cap
11
+ (first 64600 samples, shorter clips repeat-padded; the authors score whole clips). score = logit[real]
12
+ - logit[fake] (the authors'' Score_real - Score_fake on softmax probabilities is a monotone function
13
+ of it). FP32 (the authors'' eval precision; TF32 off).'
14
+ code: https://github.com/nii-yamagishilab/AntiDeepfake-AudioLLM/tree/2d445b525bb52518448b353fe1fee51d5f815af9
15
+ checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/NII-SLT26-XLSR-FC/blob/60b2822b59f136d86644e9f251092c49265569d4/nii_slt26_xlsr_fc.safetensors
16
+ params_millions: 315.439746
17
+ paper:
18
+ arxiv_id: '2610.17617'
19
+ url: https://arxiv.org/abs/2610.17617
20
+ bibtex: "@inproceedings{ge2026what,\n title = {What Do Deepfake Audio Detectors Hear? From Self-Supervised\
21
+ \ Models to Audio LLMs},\n author = {Ge, Wanying and Wang, Xin and Yamagishi, Junichi},\n booktitle\
22
+ \ = {IEEE Spoken Language Technology Workshop (SLT)},\n year = {2026}\n}\n"
23
+ dataset:
24
+ id: lab260/LRLspoof
25
+ revision: ac76f0e5472fdb68a6726e1fa94a148316e76916
26
+ split: test
27
+ scores:
28
+ srr_complement: 50.716126514278436
29
+ n_trials: 1304455
30
+ n_skipped: 286
31
+ calibration:
32
+ source_dataset: SpeechAntiSpoofingBenchmarks/DeepVoice
33
+ threshold: -7.410723315286624
34
+ artifact:
35
+ scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/NII-SLT26-XLSR-FC/resolve/5816b4e2b908d0ac19400851467267dfb63c4fd5/.eval_results/lab260/LRLspoof/scores.txt
36
+ scores_sha256: aee42a288e475430797272281d9db8cd65c22fd3ccf319cc6c1172d46e722451
37
+ bench_version: speech-spoof-bench==0.4.1
38
+ reproduction:
39
+ reproduced_by: SpeechAntiSpoofingBenchmarks
40
+ reproduced_at: '2026-11-14'
41
+ reproduced_bench_version: speech-spoof-bench==0.4.1
42
+ match: inference
43
+ submitter:
44
+ hf_username: korallll
45
+ contact: k.n.borodin@mtuci.ru
46
+ submitted_at: '2026-11-14'
47
+ notes: 'HF nii-yamagishilab/slt2026-checkpoints @ ae9d2bc72b670e49513f5bd125bf679095bf4b92, checkpoints.tar.gz
48
+ -> xlsr-fc.ckpt (sha256 0caa8335c5e5504d114f24415d20c8f0082a5b96c3b9c81dd3fe913d7db8a5bc; a torch training
49
+ checkpoint {step 79205, epoch 5, model_state_dict, optimizer_state_dict, config}; pickle globals inspected
50
+ with pickletools, loaded with weights_only=True). model_state_dict re-saved unchanged as nii_slt26_xlsr_fc.safetensors
51
+ (424 tensors, sha256 3dc2417582c6cac8287d7bdfe6fb5224de001dab19f46d38409e76dc2820c47a) + the needed
52
+ config values as JSON; re-hosted under CC BY-NC 4.0 (LICENSE-CHECKPOINT). Code BSD-3. The XLS-R encoder
53
+ is built from the facebook/wav2vec2-xls-r-300m config (HF @ 1a640f3) and every tensor comes from the
54
+ checkpoint (strict load, all 424 tensors used). Labels: real id 0, fake id 1 (checkpoint config). Verified
55
+ against the authors'' own build_encoder_fc_components + EncoderFCCollator + ClassificationHead on 18
56
+ 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
57
+ off. Deviation: Arena 4 s cap instead of whole clips (random 30 s crop if longer). Training data CompSpoof
58
+ V2 train: bona fide speech CommonVoice, LibriTTS, english-conversation-corpus; spoof speech ASVspoof
59
+ 5 and MLAAD; environment AudioCaps, TAU UAS, TUT SED, UrbanSound, VGGSound, EnvSDD, VcapAV (arXiv 2606.10791
60
+ Table II; dataset gated, splits unstated) -> LibriTTS and ASVspoof 5 taken as whole corpora -> ASVspoof5,
61
+ SONAR, CD-ADD, LibriSeVoc marked trained_on (n/a*). Every training bona fide clip is speech mixed with
62
+ an environmental sound, so clean studio speech is out of domain (most Arena clips score as fake). Self-reported
63
+ CompSpoof V2 test EER% (3-seed mean): fake event 9.50 / fake speech 21.21 / both 13.96. J-SPAW_LA EER
64
+ 36.756% (PyTorch); TensorRT FP32 on RTX 4070 Ti SUPER 36.875% (Spearman 0.99998 vs PyTorch).'