Multi-cepstral feature integration for audio deepfake detection under domain shift

Recent advances in speech synthesis and voice conversion technologies have enabled the creation of highly realistic audio deepfakes, posing significant challenges to biometric authentication, digital forensics, and misinformation detection. Although numerous countermeasures have been proposed, many existing approaches rely either on a single acoustic representation or on computationally intensive deep learning architectures, often exhibiting limited robustness when operating conditions differ from those observed during training. This paper investigates a lightweight and interpretable framework for audio deepfake detection based on the integration of multiple complementary cepstral representations. Specifically, Mel-Frequency Cepstral Coefficients (MFCCs), Linear Frequency Cepstral Coefficients (LFCCs), Constant-Q Cepstral Coefficients (CQCCs), and Gammatone Frequency Cepstral Coefficients (GFCCs) are jointly exploited to capture synthesis-related artifacts across different spectral resolutions and perceptual scales. The resulting feature representation is evaluated using conventional machine learning classifiers under strict speaker-independent protocols. Extensive experiments conducted on the FakeAVCelebV2 and ASVspoof2019 LA benchmarks, together with cross-dataset evaluations on the challenging In-The-Wild Audio (IWA) dataset, demonstrate the effectiveness of the proposed framework under both in-domain and out-of-domain conditions. Furthermore, a domain-stability analysis is introduced to investigate the contribution of individual acoustic descriptors to cross-dataset generalization and to identify features that remain robust across different data distributions. The obtained results demonstrate the effectiveness of integrating complementary cepstral representations for audio deepfake detection and highlight the importance of feature stability in improving cross-dataset generalization.