Designing Undetectable Morphing Speech: How Morph Rate and Discrete/Continuous Changes Influences Change Detection Thresholds

Altered Auditory Feedback (AAF) involves artificially modifying how speakers perceive their own voices, with potential applications in clinical treatment, emotion regulation, and self-representation. However, abrupt voice changes often draw attention and cause discomfort. This study investigates whether gradual voice conversion can avoid detection while still achieving substantial changes in vocal quality. We employed a GAN-based model (StarGANv2-VC) to interpolate a male speaker’s voice from 0% to 100% under two modes (step vs. gradual) and morph rates (0.1 vs. 0.05). Ten male participants pressed a key upon detecting any change, and their detection thresholds were measured using a Stealth Score. The gradual mode and faster morph rate yielded higher Stealth Scores, suggesting that smooth transitions and fewer silent intervals reduced detectability. These findings indicate a trade-off between step size and total duration, underscoring the need to carefully balance transformation rate and mode to develop unobtrusive AAF implementations for emotion and self-representation.