Deepfake Detection in the Generative AI Era: A Review and Proposed Audio–Transcript Framework

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Generative AI can now produce or alter convincing images, video and speech at low cost. This raises the risk of misinformation, impersonation and fraud and it weakens the reliability of recordings as evidence. This review covers how deepfakes are generated and detected across visual, audio and multimodal media, with most attention on synthetic speech. It compares acoustic representations, raw-waveform and self-supervised models and benchmark datasets and it looks at why detectors fail on unseen generators and real recordings. Recent evaluation and dataset studies report that detector performance depends heavily on channel conditions, preprocessing, attack diversity and whether training and test speakers overlap. The main gap found in this review is that transcript information has seen little use and only cautious use, as a second source of evidence for audio deepfake detection. To address it, an audio–transcript framework is proposed with three configurations that can be compared directly: audio-only, transcript-only and fused. The planned evaluation uses speaker-disjoint splits, cross-dataset tests, noise and compression stress tests, ASR-error analysis and standard anti-spoofing metrics. The framework has not been implemented, so the paper presents a research direction and an evaluation plan, not experimental results. It gives a reproducible way to test whether linguistic evidence improves robustness without mistaking unusual human language or ASR errors for signs of forgery.