A Hybrid CNN-LSTM Approach for Deepfake Audio Detection

As synthetic media techniques advance, the detection of manipulated audio content, known as deepfake audio, has become increasingly important. With the rise of challenges in cybersecurity and misinformation detection posed by deepfake audio, there is a pressing need for detection methods. This research introduces a system for recognizing deepfake audio that combines recurrent and convolutional networks and utilizes Mel Frequency Cepstral Coefficients (MFCC) for extracting features. It also explores feature extraction methods such as spectral contrast, spectral flatness and chromagram. This blended approach merges the effectiveness of MFCC in representing characteristics with CNNs expertise in capturing features and LSTMs capability in capturing temporal dependencies. The system achieves an accuracy rate of 94.73% on a combination of two datasets named ’Wave Fake’ and Release in the Wild’. By offering an adaptable solution for detecting manipulated audio content this study pushes forward the field of deepfake detection and helps tackle the challenges presented by synthetic media in audio forensics.