Quefrency Approach to Audio Deepfake Detection

This work aims at audio deepfake detection (ADD) in three different scenarios-only deepfake, re-recorded-deepfake, and replayed-deepfake. To that effect, the existing Fake or Real (FoR) dataset is enhanced by a reverberation topology leading to a novel version of the dataset (called as rev-FoR-rerec) to simulate replayed-deepfakes in varying acoustic environments. Furthermore, this work investigates the significance of quefrency-based representation for ADD. Our findings indicate that as compared to mel-spectrogram representations, the proposed quefrency-based system achieves improved EERs on both testing and validation sets of all three attack scenarios, except for the case of replayed-deepfake where the proposed system performs nearly equal to the mel-spectrogram with an absolute difference in testing EER as 0 and an absolute difference in validation EER as 0.09. Furthermore, our sub-band analysis shows that the bands 0-1000 Hz and 7000-8000 Hz are the most discriminating subbands for ADD. Experiments are also performed to investigate the effect of room size and reverberation damping factor on the detection of replayed-deepfakes.