Room Reverberation Effectively Masks Deepfake Traces

Abstract:
A study is presented on testing the performance of a publicly available AI tool (DEEPFAKE TOTAL) that is developed to analyse audio files to detect deepfakes. The audio data in the test is built on a database of brand personality speech (BRANDDB), in which actors portray brand personalities in a way that is recognisable to listeners. Deepfakes were generated from these originals by using an established voice cloning AI (ELEVENLABS). Both the originals and the deepfakes were recorded a second time while being played back in an office space in order to add real-room reverberation. All recordings were analysed for fakeness by DEEPFAKE TOTAL. The most crucial result is that deepfakes with room reverberation evaluate as not different from the originals. They even tend to be considered less fake than them, indicating that room reverberation effectively masks deepfake traces.