Deepfake Audio Detection: A Comparative Study of Advanced Deep Learning Models

The rapid advancements in artificial intelligence (AI) have significantly enhanced audio synthesis capabilities, enabling deepfake technology to replicate human speech with near-perfect accuracy. This poses severe security threats to various domains, including banking, customer service, and law enforcement, as malicious actors can exploit speech synthesis techniques such as voice conversion, replay attacks, and text-to-speech (TTS) to manipulate or impersonate individuals. Both spectral and temporal features were extracted from the audio recordings for feature extraction. This paper explores state-of-the-art detection frameworks for deepfake audio, highlighting the effectiveness of advanced deep learning (DL) frameworks. Specifically, the performance of the bidirectional long short-term memory (BLSTM) network, a custom convolutional neural network (CNN), a residual CNN integrated with an attention mechanism and bidirectional gated recurrent unit (ResCNN-Attention-BGRU), WIREnet (a variant of BLSTM), the residual network for fully connected (ResNet FC), and a squeeze-and-excitation-enhanced one-dimensional CNN (SE-Enhanced 1D-CNN) were compared. The proposed models achieved notable testing accuracies, with SE-Enhanced 1D-CNN reaching the highest at 97.64%, followed by ResNet FC (97.46%) and WIREnet (97.20%), demonstrating strong generalization across the evaluated architectures. Furthermore, we experimented with different numbers of mel-frequency cepstral coefficients (MFCCs), specifically 13, 26, and 39, in combination with other spectral and temporal features. The experimental results demonstrated that MFCC-39, together with spectral and temporal features, had a robust feature representation and achieved the best performance for deepfake detection.