Explainable fake voice detection in AI-generated audio using frequency-aware transformers

The increasing availability of AI-generated synthetic voices poses challenges to the security, privacy, and integrity of digital audio, creating a need for accurate and interpretable detection methods. This study proposes a Frequency-Aware Transformer (FAT) for detecting AI-generated fake voices by incorporating multi-band spectral embeddings, frequency-aware positional encoding, and frequency-guided attention into a transformer-based architecture. The model was evaluated on the ASVspoof 2019 Logical Access benchmark and compared with conventional deep learning, transformer-based, and self-supervised learning baselines. FAT achieved 94.6 ± 0.3% accuracy, 94.3 ± 0.3% precision, 94.7 ± 0.2% recall, 94.5 ± 0.2% F1-score, and 6.3 ± 0.2% Equal Error Rate. Compared with ECAPA-TDNN, the strongest conventional baseline, FAT improved accuracy from 91.3 to 94.6% and reduced EER from 9.7 to 6.3%. The ablation analysis showed that removing self-attention resulted in the largest accuracy reduction of 5.1% points, followed by removal of frequency encoding (3.8 points), frequency bias (2.9 points), and frequency-aware positional encoding (2.5 points), demonstrating the contribution of these components to the observed performance. Cross-dataset evaluation showed that FAT retained discriminative performance under the evaluated dataset shifts, although some degradation occurred on unseen conditions. Under controlled artificial noise corruption, FAT accuracy decreased from 94.6% in the clean condition to 88.0% at 5 dB SNR, while remaining higher than the evaluated baseline models. Attention rollout and SHAP analysis provided interpretable information about the spectral-temporal regions and frequency bands associated with model predictions. Overall, the findings indicate that incorporating frequency-domain information into transformer-based modelling can improve fake voice detection performance and provide complementary interpretability under the evaluated benchmark and acoustic perturbation conditions.