Detecção de áudios falsos através de Wavelets e Redes Neurais Convolucionais

Abstract

The advancement of Artificial Intelligence (AI) has enabled the creation of audio deepfakes— synthetic voice manipulations that mimic real human voices with high precision. This technology, while having useful and legitimate applications, represents a significant risk to information security, being also used in fraud and disinformation. The sophistication of fake audio generation algorithms makes detection a challenge, hindering the analytical capabilities of both humans and conventional algorithms, thus necessitating the development of more effective detection methods. To address this challenge, this work proposes and evaluates a methodology that combines audio signal processing and deep learning. The approach transforms audio signals into scalograms— images representing the signal’s energy in the time-frequency domain— by means of the Continuous Wavelet Transform (CWT). For a comparative analysis, the study uses scalograms generated from three wavelet families (Morlet, Complex Morlet, and Mexican Hat) to train the MobileNet Convolutional Neural Network (CNN) architecture. The experiments, conducted on the FoR dataset, demonstrate that the choice of wavelet family is a determining factor for performance. The configuration with the Morlet wavelet showed the best balanced performance, achieving an accuracy of 84.01% and an F1-Score of 84.24%, which demonstrates the feasibility of the approach and establishes this combination as the most effective in the study.