| As a research hotspot in the field of biometric identification,the speaker verification system has developed rapidly in recent decades.Because of its convenience,efficiency and low cost,it has been widely used in fields requiring identification.However,in the practical application of the speaker verification system,the existence of replay attack speech has caused more and more attention to security issues.From the perspective of application security,improving the security of the speaker verification system and increasing the ability of the speaker verification system against spoofing attacks becomes an urgent issue.In order to effectively resist replay attack speech,this paper proposes two feature extraction methods suitable for replay attack detection by analyzing the frequency response difference of the input channel,the one is Channel frequency response Difference Enhancement Cepstral Coefficient(CDECC)feature and another is Wavelet Packet Entropy Density(WPED)feature.The CDECC feature enhances the difference in channel frequency response through a third-order polynomial non-linear frequency scale transformation.The WPED feature enhances the difference in channel frequency response through wavelet packet transformation(WPT)and the entropy density values of specific wavelet packet nodes,increasing the distance between the two types of speech in the feature space.Finally,the replay attack detection module using CDECC feature and WPED feature is embedded in the speaker verification system in a tandem manner,which implements a speaker verification system with replay attack detection capability.The replay attack detection experiment based on ASVSpoof 2.0 dataset shows that the equal error rate(EER)of CDECC and WPED is 25.03%and 23.22%.The EER of CDECC and WPED has a relative reduction of 10%and 16.50%compared to the baseline system.The ASV system's false acceptance rate(FAR)is significantly reduced by embedding the replay attack detection module,and the system's EER is reduced from 4.80%to 1.07%and 1.03%. |
