File size: 1,997 Bytes
129c587 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 | 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.
|