Transformer Based Encoder Representation for Audio Anti-Spoofing: Separability and Alignment
IEEE Transactions on Audio, Speech and Language Processing · 2026 · 34 · 1674-1685
DOI: 10.1109/taslpro.2026.3671972 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Transactions on Audio, Speech and Language Processing, 2026. Verfasst von Yu-Li Ding und 2 weiteren Autor:innen. Die Arbeit nennt 47 Literaturangaben. Gefördert von: National Science and Technology Council, Innovative Center on Sustainable Negative-Carbon Resources, Featured Areas Research Center Program.
Abstract (Originalsprache)
This paper proposes a Modified Audio Anti-Spoofing method based on a pretrained transformer and the strategy of separability and alignment, which is called MAAS-SA hereafter. Traditional schemes suffer from a challenge on generalizations for cross multiple languages because their models are solely trained by a single language dataset. The MAAS-SA method utilizes three distinct language datasets to overcome the challenge mentioned above. However, a difficulty arises in effectively extracting common features among these three datasets due to the existence of high variations among them. Therefore, the MAAS-SA method comprises three main components to tackle the difficulty, which include a pretrained transformer as an encoder, a combination of the supervised contrastive, hinge, and focal losses to enhance the feature separability, and an alignment classifier with multiple boundaries based on domain generalization to promote classification ability. Moreover, the alignment classifier of the MAAS-SA method not only extracts common features among these three datasets but also identifies unique features for each dataset. Additionally, a variant of the MAAS-SA method, which is called WMAAS-SA, is developed by integrating wavelet domain signals and Kullback-Leibler divergence (KLD) loss to extract more distinguishable representations and then to improve classification ability on audio anti-spoofing. Finally, experimental results demonstrate that the WMAAS-SA method outperforms existing approaches across datasets, ASVspoof 2019 logical access, Chinese Fake Audio Detection (CFAD), French and Russian language in Multi-Language Audio Anti-Spoofing Dataset (MLAAD). Moreover, it is superior to other schemes on unseen attacks and languages.
Förderung
- National Science and Technology Council
- Innovative Center on Sustainable Negative-Carbon Resources
- Featured Areas Research Center Program
- Ministry of Education in Taiwan
Zitieren
APA
Ding, Y., Tsai, H., & Shih, Y. (2026). Transformer Based Encoder Representation for Audio Anti-Spoofing: Separability and Alignment. IEEE Transactions on Audio, Speech and Language Processing, 34, 1674-1685. https://doi.org/10.1109/taslpro.2026.3671972
Vancouver
Ding Y, Tsai H, Shih Y. Transformer Based Encoder Representation for Audio Anti-Spoofing: Separability and Alignment. IEEE Transactions on Audio, Speech and Language Processing. 2026;34:1674-1685. doi:10.1109/taslpro.2026.3671972
BibTeX
@article{Ding2026,
author = {Ding, Yu-Li and Tsai, Hung-Hsu and Shih, Yin-Tzer},
title = {Transformer Based Encoder Representation for Audio Anti-Spoofing: Separability and Alignment},
journal = {IEEE Transactions on Audio, Speech and Language Processing},
year = {2026},
volume = {34},
pages = {1674--1685},
doi = {10.1109/taslpro.2026.3671972}
}
EUVIMED