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.

The long-term window based features have been found to be effective for spoofing attack detection. One such important countermeasure is constant-Q cepstral coefficients (CQCC) that is derived from constant-Q transform. During its extraction, the octave power spectrum is converted to the linear power spectrum by performing uniform resampling. However, the information from the octave power spectrum is different from that carried by the linear power spectrum. We believe that the octave power spectrum can offer complementary information to the linear power spectrum for spoofing attack detection. In this regard, we propose to combine the coefficients generated using both linear and octave power spectrum. The combined feature is referred to as extended CQCC (eCQCC), which is hypothesized to have better discriminative information for detection of spoofing attacks. In addition, we use the short-term spectral statistics information (STSSI) along with eCQCC feature to form another novel feature representation referred to as eCQCC-STSSI to have improved anti-spoofing countermeasure. We perform the studies with the proposed features for both synthetic and playback attacks using ASVspoof 2015 and ASVspoof 2017 version 2.0 corpus, respectively. The studies reveal that eCQCC outperforms the conventional CQCC feature as well most of the known systems showing importance of octave spectrum information. Further, the hybrid feature eCQCC-STSSI improve the performance of eCQCC feature due to the STSSI information combined with it. [ABSTRACT FROM AUTHOR]