UTCN-XLSR2B-LogReg: Speech Anti-Spoofing Arena wrapper

Packaged for the Speech Anti-Spoofing Arena from the authors' code and paper.

Data-centric real-world deepfake detector (Combei et al., Interspeech 2025): frozen wav2vec 2.0 XLS-R 2B truncated to its first 9 transformer layers (the authors' released "wav2vec2-xls-r-2b_truncated"), features = hidden_states[9] of the truncated model (= output of layer 9 after the encoder's final LayerNorm, as in the authors' app.py), mean-pooled over time (1920-d), scored by a logistic regression (C=1e6) trained on the margin-pruned union of seven scientific corpora ("ALL": ASVspoof 2019 LA, FoR-norm, ASVspoof 2021 subset, TIMIT-TTS, ODSS, MLAAD v5 + M-AILABS, ASVspoof 5 train/dev), followed by the authors' MinMax score scaler. Input: 16 kHz, Arena 4 s cap (first 64600 samples, shorter clips repeat-padded; app.py scores <= 60 s segments), per-utterance zero-mean/unit-variance normalisation as the HF feature extractor. score = MinMax-scaled LR decision value (class 1 = real), higher = bona fide. FP32 (TF32 off, eager attention). Authors' released checkpoint; no license, so the weights are linked, not re-hosted.

  • Weights: not re-hosted (the authors grant no license for them). Get the authors' checkpoint; its sha256 is in the meta.yaml notes.
  • Params: 433.05 M. Trained on all splits (incl. test) of ASVspoof 2019 LA, FoR, ASVspoof 2021 DF, ODSS and others, plus ASVspoof 5 train+dev. The Arena marks ASVspoof2019_LA, ASVspoof2021_DF and ODSS as n/a* (training overlap); the released configuration was selected on In-the-Wild.
  • Score: higher = bona fide. J-SPAW_LA was scored with the PyTorch wrapper; most other datasets with the TensorRT drop-in class (FP32 on the RTX 4070 Ti SUPER; FP32 on the RTX 5060 Ti), which matched PyTorch on the full J-SPAW_LA set (|dEER| <= 0.25, Spearman >= 0.999); the RTX 4060 Ti cards ran PyTorch.

Arena results

arena tier arena rank

EER% (lower is better); EmoSpoofTTS and LRLspoof are spoof-only and use 1-SRR% at the model's DeepVoice EER threshold. Per-dataset scores are under .eval_results/.

EER% 2.43 on ASVspoof2019_LA EER% 7.92 on ASVspoof2021_LA EER% 4.08 on ASVspoof2021_DF EER% 12.68 on ASVspoof5 EER% 0.37 on InTheWild EER% 3.66 on ADD22_eval_31 EER% 12.08 on ADD2023_track12_test_r1 EER% 7.68 on CFAD EER% 1.45 on CD-ADD EER% 8.21 on CVoiceFake_small EER% 5.4 on DeepVoice EER% 1.06 on LibriSeVoc EER% 4.78 on SONAR EER% 8.07 on ArAD EER% 6.34 on DECRO EER% 0.88 on J-SPAW_LA EER% 4.86 on ODSS EER% 4.06 on HABLA EER% 11.17 on DFADD EER% 1.2 on PyAra EER% 7.54 on XMAD EER% 0.38 on EmoFake_test SRR COMPLEMENT 16.24 on EmoSpoofTTS SRR COMPLEMENT 7.28 on LRLspoof

Citation

@inproceedings{combei2025unmasking,
  title     = {Unmasking real-world audio deepfakes: A data-centric approach},
  author    = {Combei, David and Stan, Adriana and Oneata, Dan and M{\"u}ller, Nicolas and Cucu, Horia},
  booktitle = {Proc. Interspeech 2025},
  year      = {2025},
  note      = {arXiv:2506.09606}
}
Downloads last month
153
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Paper for SpeechAntiSpoofingBenchmarks/UTCN-XLSR2B-LogReg