Abstract:
Voice cloning methods have been used in a range of ways, from customized speech interfaces for marketing to video games. Current voice cloning systems are smart enough to learn speech characteristics from a few samples and produce perceptually unrecognizable speech. These systems pose new protection and privacy risks to voice-driven interfaces. Fake audio has been used for malicious purposes and is difficult to classify what is real and fake during a digital forensic investigation. This paper reviews the issue of deep-fake audio classification and evaluates the current methods of deep-fake audio detection for forensic investigation. Audio file features were extracted and visually presented using MFCC, Mel-spectrum, Chromagram, and spectrogram representations to further study the differences. Harnessing the different deep learning techniques from existing literature were compared using five iterative tests to determine the mean accuracy and the effects thereof. The results showed a Custom Architecture gave better results for the Chromagram, Spectrogram, and Me-Spectrum images and the VGG-16 architecture gave the best results for the MFCC image feature. This paper contributes to further assisting forensic investigators in differentiating between synthetic and real voices.
