Replay Attack Detection in Automatic Speaker Verification Based on ResNeWt18 with Linear Frequency Cepstral Coefficients

This paper proposes, effective method for replay attack detection used in an automatic speaker verification system. The replay attack is of interest because it is the most straightforward and effective attack and is challenging to detect. It is a playback of the recording of the voice of a target speaker. From the literature, no speech features work well with all classifiers, and there is no investigation of using ResNet-based model, called ResNeWt, with linear frequency cepstral coefficient (LFCC). Therefore, a replay attack detection model based on 18-layer ResNeWt that takes LFCCs as the input, was constructed in this paper. The proposes method was tested on a dataset provided by ASVspoof 2019 competition. In terms of the equal error rate (EER), the proposed method is the best in all existing methods, with an EER of 0.29%. The comparison in terms of replay attack detection was also made in detail. The performance of the proposed method in terms of the balanced accuracy, precision, recall, and F1-score was considerably better than existing methods.