An adaptive transmission line cochlear model based front-end for replay attack detection

Abstract

The cochlea is a remarkable spectrum analyser with desirable properties including sharp frequency tuning and level-dependent compression and the potential advantages of incorporating these characteristics in a speech processing front-end are investigated. This paper develops a framework for an active transmission line cochlear model employing adaptive notch and resonant filters. The proposed model reproduces the observed asymmetric auditory filter shape with a sharp high-frequency roll-off and level-dependent nonlinear dynamic range compression characteristics. Experimental analysis demonstrates that sharp frequency tuning and dynamic range compression of the proposed model lead to an enhanced spectral representation compared with other spectral analysis methods. The proposed model was employed in the front-end of replay spoofing attack detection systems, and experiments on the ASVspoof 2017 version 2.0 and ASVspoof 2019 databases demonstrate that the proposed model outperforms linear and nonlinear level-dependent parallel filter bank auditory models and classical spectro-temporal front-ends. The use of the proposed model leads to relative improvements of 45.6%, 51.9% and 60.8% over the baseline feature CQCCs of ASVspoof version 2.0 and CQCCs and LFCCs of ASVspoof2019 on evaluation datasets, respectively.