Audio deepfake detection using constant-Q transform scalograms and spatial-transformer attention networks

Detecting synthetically generated speech has become increasingly important for forensic speaker verification and digital media security. In particular, recent advances in speech synthesis and voice conversion technologies have enabled the generation of increasingly natural speech samples that are difficult to distinguish from genuine human speech, rendering fake speech detection approaches based solely on perceptual cues insufficient and increasing the need for more robust automated detection systems. This has highlighted the importance of developing methods capable of revealing subtle synthesis artifacts that may not be readily perceptible to the human ear in the time-frequency domain, while jointly modeling patterns across multiple scales. This study investigates a hybrid deep learning model for audio deepfake detection, using a Constant-Q Transform (CQT) scalogram frontend alongside residual convolutional blocks, spatial self-attention, and transformer encoders. This architecture is designed to capture both local spectral details and sequential patterns within audio recordings. Evaluations on the benchmark Fake-or-Real (FoR) dataset show that the proposed approach achieves an overall test accuracy of 94%, an AUROC score of 0.9919, and an Equal Error Rate (EER) of 3.86%, with a bootstrap analysis 95% confidence interval: 93.50–95.95%). These results demonstrate that combining time-frequency representations with attention mechanisms provides a stable decision framework for identifying speech spoofing attacks under standard evaluation protocols. The findings further demonstrate that, rather than relying on approaches based on a single feature extraction mechanism, hybrid architectures that jointly capture local spectral characteristics, attention-based regional information, and long-range temporal dependencies hold significant potential for audio deepfake detection. In this regard, the study provides a basis for the development of reliable and generalizable detection systems against increasingly sophisticated synthetic speech generation methods.