Leveraging Acoustic Features and Deep Neural Architectures for Audio Deepfake Detection

This research presents a comparative analysis of various audio features and high-level architectures for deefake detection with emphasis on computational efficiency. Several light-weight models are proposed as opposed to GAN-based approaches in literature for evaluating custom-generated deep-fakes. The model is trained on Fake or Real dataset and achieved commendable performance using MFCC-Conformer and MFCC-LSTM feature-model combinations by achieving 87.61% and 87.52% accuracy, respectively. Specifically, the MFCC-Conformer recorded a TN of 526 and a FN of 18, along with an AUC score of 0.96, while the MFCC-DenseNet achieved a TN of 535, an FN of 9, and an AUC score of 0.96, underscoring their effectiveness in identifying fake audios. The outcomes underscore the effectiveness of the proposed models in combating the proliferation of misleading media content.