Listening Between The Lines : Investigating The Correlation Between Voice Recognition and Audio Deepfake Differentiation

Sammanfattning

Deepfake technology is a rapidly evolving field that causes some concerns of disinformation spreading within the political sphere. Many studies have been conducted on the abilities of both human and AI tools to detect deepfake videos, audio and images. However, audio deepfakes within the context of Swedish politics remain largely unexplored. This study investigates whether Swedish-speaking individuals who can identify Swedish politicians have a higher accuracy in detecting audio deepfakes of those politicians. Although some of our participants used their existing impression of what the politician usually sounds like to determine whether the audio samples were genuine or not, our overall results imply a weak positive correlation between recognition ability and deepfake detection ability. This means that participants who can identify the voice of a person are slightly more accurate at distinguishing deepfake samples and real voice samples of that person. However, the weak positive correlation was not statistically significant.