An Adaptive-Q Cochlear Model for Replay Spoofing Detection

Tharshini Gunendradasan, Eliathamby Ambikairajah, Julien Epps, Haizhou Li

Replay attack poses a key threat for automatic speaker verification systems. Spoofing detection systems inspired by auditory perception have shown promise to date, however some aspects of auditory processing have not been investigated in this context. In this paper, a transmission line cochlear model that incorporates an active feedback mechanism is proposed for replay attack detection. This model compresses the considerable energy variation in each auditory sub-band filter by boosting low-amplitude signal, an effect that is not considered in many auditory models. To perform the compression, the parameters of each auditory sub-band filter are modified based on the sub-band energy, analogous to the effect of the closed-loop adaptation mechanism that allows perception of a wide dynamic range from a physically constrained system, which we term adaptive-Q. Evaluation on the ASVspoof 2017 version 2 database suggests that the adaptive-Q compression provided by the proposed model helps to improve the performance of replay detection, and a relative reduction in EER of 26% was achieved compared with the best results reported for auditory system-based feature proposed for replay attack detection.

 DOI: 10.21437/Interspeech.2019-2361

Cite as: Gunendradasan, T., Ambikairajah, E., Epps, J., Li, H. (2019) An Adaptive-Q Cochlear Model for Replay Spoofing Detection. Proc. Interspeech 2019, 2918-2922, DOI: 10.21437/Interspeech.2019-2361.

  author={Tharshini Gunendradasan and Eliathamby Ambikairajah and Julien Epps and Haizhou Li},
  title={{An Adaptive-Q Cochlear Model for Replay Spoofing Detection}},
  booktitle={Proc. Interspeech 2019},