Deepfake Audio Detection in Voice Authentication: A Spectral and CNN-Based Comprehensive Review

As voice authentication systems become increasingly integral to critical domains such as banking, smart assistants, and remote identity verification, they face escalating threats from AI-generated audio, commonly referred to as deepfakes. These synthetic voices, produced through advanced text-to-speech and voice conversion technologies, can convincingly imitate human speech, thereby undermining the reliability and security of authentication frameworks. This study provides a comprehensive review of spectral-based techniques for deepfake audio detection, highlighting the roles of spectrograms, Mel-Frequency Cepstral Coefficients (MFCC), and Constant-Q Transform (CQT) in exposing time-frequency anomalies. The integration of Convolutional Neural Network (CNN)-based spoof detection modules before identity verification is identified as a critical architectural strategy to enhance system resilience. This review also outlines the prevailing challenges, including vulnerability due to emerging generative models, limited interpretability of deep learning classifiers, and decreased robustness under realistic or noisy conditions. To advance the field, this study emphasizes promising research directions such as hybrid modeling approaches, adversarial training techniques, and the development of multilingual open-access deepfake audio datasets. By critically synthesizing existing research, this review aims to inform the design of more robust, generalizable, and transparent voice authentication systems capable of surviving the evolving landscape of audio-based threats.