Codecfake-W2V2-AASIST-CoTrain: Speech Anti-Spoofing Arena wrapper

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

W2V2-AASIST countermeasure from the Codecfake paper: a frozen wav2vec 2.0 XLS-R 300M front-end (per-utterance mean/variance normalised input; hidden states after its 5th transformer layer, i.e. HF hidden_states[5]) feeding an AASIST head (linear 1024->128, residual encoder, spectro-temporal graph attention). Weights: the co-trained release (ASVspoof2019 LA train + Codecfake train, with the paper's CSAM sampler; the paper's headline countermeasure). The upstream repo carries no license, so the weights are linked from the authors' GitHub repo (pinned commit), not re-hosted. FP32, first 64600 samples (repeat-tiled if shorter); score = log-odds of bona fide.

• Weights: not re-hosted: the upstream repo carries no license. The wrapper downloads the authors' checkpoint from the pinned commit and checks its sha256.
• Params: 76.554634 M. Trained on ASVspoof 2019 LA train + Codecfake (Xie et al.) train.
• 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{xie2024codecfake,
  title         = {The Codecfake Dataset and Countermeasures for the Universally Detection of Deepfake Audio},
  author        = {Xie, Yuankun and Lu, Yi and Fu, Ruibo and Wen, Zhengqi and Wang, Zhiyong and Tao, Jianhua and Qi, Xin and Wang, Xiaopeng and Liu, Yukun and Cheng, Haonan and Ye, Long and Sun, Yi},
  journal       = {arXiv preprint arXiv:2405.04880},
  year          = {2024},
  eprint        = {2405.04880},
  archivePrefix = {arXiv}
}