Synthetic Speech Detection Using Extended Constant-Q Symmetric-Subband Cepstrum Coefficients and CBAM-ResNet
Yaqin Zhao, Qinyu Ding, Longwen Wu, Ruchen Lv, Jiepeng Du, Shenyang He · 2023
Synthetic speech detection has always played a considerable role in the security of automated speaker verification systems. This work proposes an extended constant-Q symmetric-subbandd cepstrum coefficients (eCQSCC) feature with phase-symbol magnitude phase spectrum information (PMPS). The eCQSCC symmetrically divides the octave spectrum and extends the dynamic information of the octave spectrum and the linear PMPS. This division and extended information make eCQSCC more capable of deceptive artifact extraction of synthetic speech. Then we improved the residual network as the back-end classifier using the convolutional block attention module (CBAM). The network proposed consists of thirty-three basic residual blocks and two CBAMs. Finally, we fed eCQSCC into this network for synthetic speech detection. Experiments on the ASVspoof2019 synthetic speech dataset show that the EER and t-DCF of this method are 0.04% and 0.001, respectively, which are significantly better than many recently known methods.