Instantaneous Phase and Excitation Source Features for Detection of Replay Attacks

In present era, the spoof detection has become an integral part of biometric systems and speaker verification is no exception to it. The replay attacks are the most common, where the attacker plays the recorded speech of a user to validate a false identity claim. Currently, constant-Q cepstral coefficient (CQCC) feature based system represents the standalone benchmark for spoof detection. However, we hypothesize that the phase and excitation source information of speech may carry additional artifacts that are useful for identifying the replay attacks. In this regard, instantaneous frequency cosine coefficients and two source features namely, discrete cosine transform of integrated linear prediction residual and residual mel frequency cepstral coefficient are explored. The studies are conducted on ASVspoof 2017 Version 2.0 database designed for the replay attacks. The results reveal that the phase and source features although perform poorer than CQCC, their fusion helps to achieve an improved performance. This indicates the complementary nature of information carried by the stated features is useful for detecting replay attacks. Further, an analysis on the behavior of each of these features under different replay configurations is also presented to highlight their effect in different scenarios.