Replay Attack Detection in Automatic Speaker Verification Using Gammatone Cepstral Coefficients and ResNet-Based Model

抄録

This paper presents a method for detecting replay attacks in an automatic speaker verification system. The replay attack is of interest because it is the most straightforward, effective attack, and difficult to detect. Even though many speech features and classifiers have been proposed, the detection performance, such as an equal error rate (EER), accuracy, and balanced accuracy, need to be improved. Therefore, we propose a method for replay attack detection that applies the Gammatone cepstral coefficients with a ResNet-based model. The proposed method was evaluated and compared with existing methods and baselines in the ASVspoof 2019 challenge. The results indicated that the proposed method outperforms our previous method and the baselines in which the EER was 8.4%. In addition, the accuracy and balanced accuracy of the spoofing detection were improved.