Detecting Synthesized Audio Files Using Graph Neural Networks

Abstract

The problem of generalization of multimodal data in the detection of artificially synthesized audio files is studied. As a solution to the problem, a method is proposed that combines a one-time analysis of the characteristics of an audio file and its semantic component, presented in the form of text. The approach is based on graph neural networks and algorithmic approaches based on keyword and text sentiment analysis. The conducted experimental studies confirmed the validity and effectiveness of the proposed approach.