Eroding Trust in Real Speech: A Large-Scale Study of Human Audio Deepfake Perception
A recent study investigates the impact of audio deepfakes on human trust in real speech. The research, which involved over 35,000 judgments from nearly 1,800 participants, reveals a significant decline in the accuracy of identifying genuine audio samples. The findings suggest that the primary concern with modern deepfakes is not just deception, but the erosion of trust in authentic speech.
- ▪The study collected 35,532 judgments from 1,768 participants.
- ▪Accuracy in identifying real audio samples dropped from 72.7% to 64.1%.
- ▪Participants are increasingly skeptical of authentic speech rather than worse at detecting deepfakes.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Sound arXiv:2605.26136 (cs) [Submitted on 21 May 2026] Title:Eroding Trust in Real Speech: A Large-Scale Study of Human Audio Deepfake Perception Authors:Nicolas M. Müller, Wei Herng Choong View a PDF of the paper titled Eroding Trust in Real Speech: A Large-Scale Study of Human Audio Deepfake Perception, by Nicolas M. M\"uller and 1 other authors View PDF HTML (experimental) Abstract:Audio deepfakes have improved rapidly recently, yet their effect on human trust in real speech remains unstudied. We present the largest listening study on audio deepfake perception to date, collecting 35,532 judgments from 1,768 participants across 138 text-to-speech and voice conversion systems.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
