Preserving Integrity: A Binary Classification Approach to Unmasking Artificially Generated Voices in the Age of Deepkakes

Generative AI uses machine learning techniques like semi-supervised or unsupervised learning algorithms for the creation of digital content such as images, audio, and videos. There are ethical concerns arising from generative AI's impact on speech technology, specifically, voice cloning and real-time voice conversion. To mitigate the associated risks of privacy breaches and misrepresentation, the research utilizes a DEEP-VOICE dataset that comprises audio clips from eight notable figures, converted using Retrieval-based Voice Conversion to detect Deepfake audio files. Framed as a binary classification problem, statistical analysis reveals significantly different distributions of temporal audio features between real and AI-generated speech. The experimental results show that the Random Forest Classifier over the 5-fold Cross-Validation technique results in a classification accuracy of 98.574%.