XLSR-Nes2Net-LA: Speech Anti-Spoofing Arena wrapper

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

XLSR-Nes2Net-LA (Zhang et al., MultiAPI Spoof, 2025): a wav2vec 2.0 XLS-R 300M front-end fine-tuned end-to-end with a Nes2Net-X (nested Res2Net TDNN) back-end extended by a local-window cross-block attention (window 3, single head, learned block position embedding) between the nested blocks, followed by mean temporal pooling and a 2-way linear classifier. Trained on a multi-corpus mix (TIMIT, ODSS, FoR, AI4T, ASVspoof 5, MLAAD and MultiAPI-Spoof). Authors' released epoch_5_1.422.pth checkpoint, FP32, deterministic first-64600-sample window (repeat-padded if shorter), no normalisation. score = logit of the bona fide class.

• Weights: not re-hosted (the authors grant no license for them). Get the authors' checkpoint; its sha256 is in the meta.yaml notes.
• Params: 317.969191 M. Trained on TIMIT, ODSS, FakeOrReal, AI4T, ASVspoof 5 train, MLAAD and MultiAPI. The Arena marks ODSS as n/a* (training overlap). The checkpoint was selected on In-the-Wild; that cell is ranked normally (maintainer's choice).
• 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

@misc{zhang2025multiapispoofmultiapidataset,
      title={MultiAPI Spoof: A Multi-API Dataset and Local-Attention Network for Speech Anti-spoofing Detection},
      author={Xueping Zhang and Zhenshan Zhang and Yechen Wang and Linxi Li and Liwei Jin and Ming Li},
      year={2025},
      eprint={2512.07352},
      archivePrefix={arXiv},
      primaryClass={cs.SD},
      url={https://arxiv.org/abs/2512.07352},
}