Detecting fake audio using convolutional neural networks for reducing misinformation

Abstract

Recently, a proliferation of techniques capable of replicating sounds has emerged, encompassing manipulated recordings, synthesized audio, and deepfake technologies. These advanced methods for generating artificial sounds have inadvertently given rise to a multitude of issues, prominently including the dissemination of misinformation, propaganda, and significant reputational damage. This article addresses this critical problem by proposing the application of a Convolutional Neural Network (CNN) model designed to predict the authenticity of a given sound. For feature extraction from the audio, Mel Frequency Cepstral Coefficients (MFCC) are utilized. To rigorously assess the robustness of the proposed model, the intricate combinations of these extracted features were further analyzed using derivatives of the MFCC features. The Fake-or-Real (FoR) dataset was used to train and test the proposed model. The Accuracy, F1-score, Precision, Recall, and Loss are used as metrics to evaluate the model’s performance. The model performed well, with an accuracy of 99.64%. The proposed model was comprehensively evaluated by systematically varying the extracted features through different derivative orders, where the accuracy decreased to 94.34%. The experimental results demonstrate the effectiveness of the model in accurately detecting fabricated audio.