SpoofSUPERB-XLSR: Speech Anti-Spoofing Arena wrapper

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

Spoof-SUPERB linear-head detector (Ali et al., 2026) over a frozen XLS-R 300M (wav2vec 2.0, 24 layers, 1024-d; s3prl xls_r_300m): the s3prl Featurizer softmax-weights all 25 hidden states (24 transformer-layer inputs + the encoder output; no per-layer norm), then Linear(1024,256) -> mean pooling over frames -> Linear(256,2). Only the layer weights, projector and classifier are trained (ASVspoof 2019 LA train, 50 epochs, SWA); the authors' released head with the s3prl converted upstream. Input as the authors' scoring code: 16 kHz, first 64600 samples (shorter clips tiled), the upstream's own per-utterance waveform layer-norm; FP32. score = logits[:,1] (bona fide).

• Weights: re-hosted here under the upstream license (xls_r_300m.pth, commit-pinned); upstream: https://github.com/issflab/spoof_SUPERB/tree/8b457584e4120ba525f720e714d0a4b2ed400d68
• Params: 317.653531 M. Trained on ASVspoof 2019 LA train (frozen XLS-R 300M; trained weighted-sum + linear head).
• Score: higher = bona fide. J-SPAW_LA was scored with the PyTorch wrapper; the 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.

Arena results

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/.

Citation

@article{ali2026superbstyle,
  title={A SUPERB-Style Benchmark of Self-Supervised Speech Models for Audio Deepfake Detection},
  author={Ali, Hashim and Adupa, Nithin Sai and Subramani, Surya and Malik, Hafiz},
  journal={arXiv preprint arXiv:2603.01482},
  year={2026}
}