CQT-ViT: Vision Transformer-Based Deepfake Audio Detection Using Constant-Q Spectrograms

The growing accessibility of deepfake audio technologies has intensified concerns about audio forgery and digital deception. This study investigates a novel approach for deepfake audio detection by leveraging Constant-Q Transform (CQT) spectrograms in conjunction with Vision Transformer (ViT) architectures. Audio signals are transformed into spectral images using CQT, capturing both frequency and temporal cues, which are then processed by ViT to identify synthesized audio patterns. Rigorous evaluation across a benchmark dataset demonstrates that the proposed CQT-ViT method attains a remarkable accuracy of 95.63%, with a precision of 96.96%, recall of 95.29%, and an F1-score of 0.96. These metrics underline the effectiveness of integrating advanced signal representation techniques with vision-based deep learning models for audio forensic tasks. The results suggest a promising direction for robust, scalable detection of artificially generated speech—a necessity as synthetic media grows in sophistication.