system:
  name: "Nes2Net"
  slug: "nes2net"
  description: >
    wav2vec 2.0 (XLS-R 300M) self-supervised front-end fine-tuned end-to-end with
    a Nes2Net-X (Nested Res2Net TDNN) back-end for speech anti-spoofing. The
    nested Res2Net structure couples multi-scale residual groups with squeeze-
    excitation, replacing dimensionality-reducing necks; mean temporal pooling +
    linear classifier. Only ~0.51M back-end params. Official Nes2Net-X single
    checkpoint (ASVspoof2021 LA 1.73% / DF 1.65% EER as reported), trained on
    ASVspoof2019 LA with RawBoost, FP32, deterministic first-64600-sample window
    (no random crop).
  code: "https://github.com/Liu-Tianchi/Nes2Net_ASVspoof_ITW"
  checkpoint: "https://huggingface.co/SpeechAntiSpoofingBenchmarks/Nes2Net"
  params_millions: 317.9026
  paper:
    arxiv_id: "2504.05657"
    url: "https://arxiv.org/abs/2504.05657"
    bibtex: |
      @article{Nes2Net,
        author={Liu, Tianchi and Truong, Duc-Tuan and Das, Rohan Kumar and Lee, Kong Aik and Li, Haizhou},
        journal={IEEE Transactions on Information Forensics and Security},
        title={Nes2Net: A Lightweight Nested Architecture for Foundation Model Driven Speech Anti-Spoofing},
        year={2025},
        volume={20},
        pages={12005--12018},
        doi={10.1109/TIFS.2025.3626963}
      }
notes: >
  XLS-R 300M (wav2vec 2.0) front-end + Nes2Net-X back-end, the single (non-averaged)
  checkpoint from Liu-Tianchi/Nes2Net_ASVspoof_ITW (Nes_ratio [8,8], SE_ratio [1],
  pool_func 'mean', dilation 2). Architecture is built from the base xlsr2_300m.pt
  model config, then every weight is overwritten by the fine-tuned checkpoint.
  Deterministic first-64600-sample window (no random crop), matching the source
  data_utils_SSL.py::pad used at eval (default --test_protocol 4sec). score = output
  logit for class 1 (bona fide); higher = more bona fide. Back-end params ~0.51M;
  params_millions reports the full deployed model incl. the XLS-R front-end.