Add NII-SLT26-Whisper-FC submission for PyAra (EER=53.6406)

#63
by korallll - opened
submissions/nii-slt26-whisper-fc.yaml ADDED
@@ -0,0 +1,62 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ schema_version: 4
2
+ system:
3
+ name: NII-SLT26-Whisper-FC
4
+ slug: nii-slt26-whisper-fc
5
+ description: 'Whisper + FC from Ge, Wang and Yamagishi (IEEE SLT 2026): a fully fine-tuned Whisper-large-v2
6
+ encoder (transformers WhisperEncoder; the decoder is never used), mean pooling of the last hidden
7
+ state over all 1500 frames, and a Linear(1280, 2) head [real, fake]. Trained on CompSpoof V2 (ESDD2)
8
+ train with the "any component fake = fake" labels, the unmixed "original" class excluded (protocol
9
+ AnyFakeNoOriginal), seed 2, best-validation-loss checkpoint. Authors'' released checkpoint. Input
10
+ as the authors'' collator: 16 kHz waveform, Arena 4 s cap (first 64600 samples, shorter clips repeat-padded;
11
+ the authors score whole clips), then zero-padded to Whisper''s 30 s and converted to the 80-bin log-mel
12
+ of WhisperFeatureExtractor (no amplitude normalisation); the mean pool includes the padded frames,
13
+ as in training. score = logit[real] - logit[fake] (the authors'' Score_real - Score_fake on softmax
14
+ probabilities is a monotone function of it). FP32 (the authors'' eval precision; TF32 off).'
15
+ code: https://github.com/nii-yamagishilab/AntiDeepfake-AudioLLM/tree/2d445b525bb52518448b353fe1fee51d5f815af9
16
+ checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/NII-SLT26-Whisper-FC/blob/aed40234b7a636287238c971aa934b62659fed28/nii_slt26_whisper_fc.safetensors
17
+ params_millions: 636.787202
18
+ paper:
19
+ arxiv_id: '2610.17617'
20
+ url: https://arxiv.org/abs/2610.17617
21
+ bibtex: "@inproceedings{ge2026what,\n title = {What Do Deepfake Audio Detectors Hear? From Self-Supervised\
22
+ \ Models to Audio LLMs},\n author = {Ge, Wanying and Wang, Xin and Yamagishi, Junichi},\n booktitle\
23
+ \ = {IEEE Spoken Language Technology Workshop (SLT)},\n year = {2026}\n}\n"
24
+ dataset:
25
+ id: SpeechAntiSpoofingBenchmarks/PyAra
26
+ revision: 40167badd6d8a289f02054761ee5411f49227b87
27
+ split: test
28
+ scores:
29
+ eer_percent: 53.64056571515673
30
+ n_trials: 201778
31
+ n_skipped: 0
32
+ artifact:
33
+ scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/NII-SLT26-Whisper-FC/resolve/5c2e359236c5b9c33e5446a52735b2b2f84b1166/.eval_results/SpeechAntiSpoofingBenchmarks/PyAra/scores.txt
34
+ scores_sha256: ca563c6f39b927f6a3574fe647b900e3a7334d853f5b3f9a8e50db1130b2d841
35
+ bench_version: speech-spoof-bench==0.4.1
36
+ reproduction: {}
37
+ submitter:
38
+ hf_username: korallll
39
+ contact: k.n.borodin@mtuci.ru
40
+ submitted_at: '2026-11-10'
41
+ notes: 'HF nii-yamagishilab/slt2026-checkpoints @ ae9d2bc72b670e49513f5bd125bf679095bf4b92, checkpoints.tar.gz
42
+ -> whisper-fc.ckpt (sha256 2c7299ffcdb749e4d162efc49f3f7acd42bde6ee7dc9c76f6c04635643b8c858; a torch
43
+ training checkpoint {step 15600, epoch 1, model_state_dict, optimizer_state_dict, config}; pickle globals
44
+ inspected with pickletools, loaded with weights_only=True). model_state_dict minus the 772 encoder.model.decoder.*
45
+ tensors (the Whisper decoder: never called by the encoder+FC model, no optimizer state, i.e. untouched
46
+ openai/whisper-large-v2 weights) re-saved unchanged as nii_slt26_whisper_fc.safetensors (489 tensors,
47
+ sha256 f3e640703416f60d5afab7a1463e7f918471aa9e6bf4b76be193087713d0ed9a) + the needed config values
48
+ as JSON; re-hosted under CC BY-NC 4.0 (LICENSE-CHECKPOINT). Code BSD-3. The encoder is built from the
49
+ openai/whisper-large-v2 config (HF @ ae46427) and every tensor comes from the checkpoint (strict load,
50
+ all 489 used). Labels: real id 0, fake id 1. Log-mel = WhisperFeatureExtractor''s torch path (run on
51
+ the GPU, <1e-5 from the CPU extractor). Verified against the authors'' own build_encoder_fc_components
52
+ + EncoderFCCollator (WhisperProcessor, padding to 30 s) + ClassificationHead on 18 J-SPAW_LA clips:
53
+ max |d logit-diff| 4.8e-6 with the 4 s cap (their batch 1 and 8), 3.8e-6 with the window off. Deviation:
54
+ Arena 4 s cap instead of whole clips (random 30 s crop if longer); the 30 s zero padding is kept (model
55
+ input contract). Training data CompSpoof V2 train: bona fide speech CommonVoice, LibriTTS, english-conversation-corpus;
56
+ spoof speech ASVspoof 5 and MLAAD; environment AudioCaps, TAU UAS, TUT SED, UrbanSound, VGGSound, EnvSDD,
57
+ VcapAV (arXiv 2606.10791 Table II; dataset gated, splits unstated) -> LibriTTS and ASVspoof 5 taken
58
+ as whole corpora -> ASVspoof5, SONAR, CD-ADD, LibriSeVoc marked trained_on (n/a*). Every training bona
59
+ fide clip is speech mixed with an environmental sound, so clean studio speech is out of domain. Self-reported
60
+ CompSpoof V2 test EER% (3-seed mean): fake event 8.29 / fake speech 23.15 / both 13.89. J-SPAW_LA EER
61
+ 14.652% (PyTorch FP32); TensorRT on RTX 4070 Ti SUPER runs FP16: 14.602% (Spearman 0.999887 vs PyTorch;
62
+ FP32 engine identical to PyTorch).'