A Multi-Domain Feature Framework for Robust Deepfake Audio Detection

Resumen

Audio deepfakes generated by modern text-to-speech and voice conversion systems pose serious threats to security, privacy, and trust in digital communication. This study proposes a multi-domain feature fusion framework for robust deepfake audio detection under realistic, in-the-wild conditions. A large-scale dataset comprising 31,780 audio samples, evenly split between genuine and synthetic speech and covering diverse speakers, languages, and recording environments, is utilized. Acoustic, compression-related, emotional, phase-based, prosodic, and statistical–spectral features are extracted and fused, and classification is performed using a lightweight fully connected neural network evaluated via stratified five-fold cross-validation. The proposed system achieves an average validation accuracy of 98.03% and an AUC of 0.998, demonstrating strong and stable discriminative performance. Ablation experiments and SHAP-based analysis highlight the critical role of compression-related features in enhancing robustness when combined with other feature domains. While the study focuses on audio-only detection and does not address adversarial or multimodal scenarios, the results indicate that multi-domain feature fusion offers a practical and generalizable solution for real-world deepfake audio detection, particularly in environments involving diverse codecs and synthesis techniques.