Detection of Replay-Spoofing Attacks Using Frequency Modulation Features

Prevention of malicious spoofing attacks is currently acknowledged as a priority area of investigation for the deployment of automatic speaker verification systems. Various features of speech signals have been used to fight counterfeit attacks. Among the different spoofing attack variants, replay attacks pose a significant threat as they do not require any expert knowledge and are difficult to detect. This paper proposes the use of a spectral centroid based frequency modulation (FM) features that we term spectral centroid deviation (SCD) for replay attack detection. Spectral centroid frequency (SCF) and spectral centroid magnitude coefficient (SCMC) features extracted from the same front-end as SCD are also investigated as complementary features. Evaluations on the ASVspoof 2017 dataset indicate that the proposed SCD features with a Gaussian Mixture Model (GMM) back-end is highly capable of discriminating genuine from replay spoofed speech, providing an equal error rate improvement greater than 60% relative to the CQCC baseline system from the ASVspoof 2017 challenge. Interestingly, experiments also reveal that the proposed SCD features exhibit an increased variance for replay spoofed speech relative to genuine speech, particularly for the lowest and highest frequency subbands.