Anti-Spoofing Speaker Verification System with Multi-Feature Integration and Multi-Task Learning

Rongjin Li, Miao Zhao, Zheng Li, Lin Li, Qingyang Hong


Speaker anti-spoofing is crucial to prevent security breaches when the speaker verification systems encounter the spoofed attacks from the advanced speech synthesis algorithms and high fidelity replay devices. In this paper, we propose a framework based on multiple features integration and multi-task learning (MFMT) for improving anti-spoofing performance. It is important to integrate the complementary information of multiple spectral features within the network, such as MFCC, CQCC, Fbank, etc., as often a single kind of feature is not enough to grasp the global spoofing cues and it generalizes poorly. Furthermore, we propose a helpful butterfly unit (BU) for multi-task learning to propagate the shared representations between the binary decision task and the other auxiliary task. The BU can obtain task representations of other branch during forward propagation and prevent the gradient from assimilating the branch during back propagation. Our proposed system yielded an EER of 9.01% on ASVspoof 2017, while the best single system and the average scores fusion obtained the evaluation EER of 2.39% and 0.96% on ASVspoof 2019 PA, respectively.


 DOI: 10.21437/Interspeech.2019-1698

Cite as: Li, R., Zhao, M., Li, Z., Li, L., Hong, Q. (2019) Anti-Spoofing Speaker Verification System with Multi-Feature Integration and Multi-Task Learning. Proc. Interspeech 2019, 1048-1052, DOI: 10.21437/Interspeech.2019-1698.


@inproceedings{Li2019,
  author={Rongjin Li and Miao Zhao and Zheng Li and Lin Li and Qingyang Hong},
  title={{Anti-Spoofing Speaker Verification System with Multi-Feature Integration and Multi-Task Learning}},
  year=2019,
  booktitle={Proc. Interspeech 2019},
  pages={1048--1052},
  doi={10.21437/Interspeech.2019-1698},
  url={http://dx.doi.org/10.21437/Interspeech.2019-1698}
}