This paper addresses the limited accuracy of existing automatic systems for detecting deepfake audio content in real time. A solution is proposed to increase the efficiency of detecting signs of deepfake use by improving neural network models and algorithms for analyzing audio recordings of human voices. An algorithm and corresponding software for a voice attack detection system have been developed. For training and testing, datasets were created containing real voice audio recordings and deepfake audio samples. Evaluation on a real-world test set demonstrated an accuracy rate of 83%, confirming the effectiveness and practical applicability of the proposed solution in combating audio deepfake threats.

Keywords: deepfake, speech audio signal, machine learning models, convolutional neural network, vishing