Enhancing Audio Deepfake Detection using Support Vector Machines and Mel-Frequency Cepstral Coefficients

Abstract

This paper presents a machine learning system designed to differentiate real from synthetic speech using a Support Vector Machine (SVM) classifier. Trained on the 'for-original' Fake-or-Real (FoR) dataset, which consists of over 195,000 genuine and computer-generated utterances, the system uses Mel Frequency Cepstral Coefficients (MFCCs) to extract features. Evaluation results show a promising accuracy of 97.28%, indicating the system's potential efficacy in real-world applications. The work lays the foundation for future improvements in detection robustness and reliability by highlighting the significance of raw data in classifier training for deepfake detection.