Audio deepfake detection has become a crucial area due to the fast development of Text-to-Speech (TTS) and Voice Conversion (VC) technologies. While these technologies improve user experiences in navigation and robotics, they also create serious security risks if misused for fraud or disinformation. This paper provides a systematic approach for detecting AI-generated audio by analysing various discriminative features, including Mel Frequency Cepstral Coefficients (MFCC), Linear Frequency Cepstral Coefficients (LFCC), and Constant Q Cepstral Coefficients (CQCC). We evaluate performance using DSP algorithms, STFT, DCT, and FIR Filter. The results emphasise the ongoing challenge of model generalisation in real-world situations. We examine the effects of pre-emphasis, Hamming windowing, and the Discrete Cosine Transform (DCT) on the visibility of synthetic data.
Prapti Kapoor (Thu,) studied this question.