Search

Conventional speaker verification systems become frail or incompetent while facing attack from spoofed speech. Presently many anti-spoofing countermeasures have been studied for automatic speaker verification. It has been known that the salient feature is of a more important role rather than the selection of classifiers in the current research field of spoofing detection. The effectiveness of constant-Q transform (CQT) has been demonstrated for anti-spoofing feature analysis in many research literatures on automatic speaker verification. On the basis of CQT-based information sub-features, i.e. octave-band principal information (OPI), full-band principal information (FPI), short-term spectral statistics information (STSSI) and magnitude-phase energy information (MPEI), three concatenated features are proposed by investigating their information complementarity in this paper, the first one is constant-Q statistics-plus-principal information coefficients (CQSPIC) by combining OPI, FPI and STSSI; the second one is constant-Q energy-plus-principal information coefficients (CQEPIC) by combining OPI, FPI and MPEI and the third one is constant-Q energy-statistics-principal information coefficients (CESPIC) by combining OPI, FPI, MPEI and STSSI. In this paper, we set up deep neural network (DNN) classifiers for evaluation of the proposed features. Experiments show that the proposed features can outperform some commonly used features meanwhile the proposed systems give better or comparable performance comparing with state-of-the-art performance on ASVspoof 2019 logical access and physical access corpus.