Deepfake technology has been driven by advanced machine learning and revolutionized multimedia creation by synthesizing hyper-realistic content. It includes images, videos, and audio. While its creative applications in entertainment and accessibility are significant, the technology also poses critical risks, especially in fraud, disinformation, and identity theft. Audio deepfakes are a subset of this phenomenon that replicate human voices with enhanced precision, mimicking tone, accent, and subtle vocal nuances. This has raised concerns in security-sensitive domains like voice authentication and forensic investigations. This systematic literature review (SLR) adopts PRISMA guidelines to explore the state-of-the-art in audio deepfake detection. It examines existing methodologies, features, datasets, and evaluation metrics across 27 studies. The review identifies traditional feature-based techniques like MFCC and LFCC, which, while effective, are limited by their dependency on manual engineering. In comparison, advanced frameworks, including CNNs, transformer-based architectures, and multimodal approaches, have achieved superior performance but often lack generalization in real-world scenarios. The findings highlight significant challenges, including dataset biases, adversarial vulnerabilities, and noise sensitivity, which limit scalability. Datasets such as FakeAVCeleb and DFDC achieved high accuracies but revealed performance inconsistencies due to dataset-specific characteristics. The review highlights the need for a generalizable audio deep fake detection framework that can achieve high accuracy across datasets.
Alnaqbi et al. (Fri,) studied this question.