Speaker recognition anti-spoofing using linear prediction residual

Speaker recognition systems have recently proven to be highly vulnerable against spoofing attacks. Therefore, it is important to detect spoofing attacks performed by speech synthesis and voice conversion in order to improve the reliability of speaker recgnition systems. To this end, in this study, we propose to use of features extracted from linear prediction (LP) resdiual signal for the detection of SS and VC based spoofing attacks against speaker recognition systems. Experiments are conducted on recently released ASVspoof 2015 database which consists of spoofed speech signals generated by ten different SS and VC algorithms. Experimental results show that, SS and VC attacks can effectively be detected using the features extracted from LP residual signal and mel frequency cepstral coefficients (MFCC) using Gaussian mixture model (GMM) classifier.