Detection of Replay Attack Based on Normalized Constant Q Cepstral Feature

Since the voice is easy to be recorded and replayed, the replay attack is considered a major threat to the voiceprint authentication system. However, few works have focused on the text-independent detection. We found that there exist differences in spectral features between the original voice and the replayed voice. Hence, constant Q cepstral coefficients which can well describe the spectral features are extracted as acoustic features. Then the cepstral mean and variance normalization is used as the postprocessing method to eliminate the impact of channel noise on detection performance. Finally, the Gaussian Mixture Model determines whether the suspected voice is replayed or not. The experimental results on the ASVspoof 2017 database indicate the proposed algorithm can significantly reduce the equal error rate and keep high robustness when the replayed voice came from various spoofing devices.