Skip to content

Voice "cloning" is actually style transfer

Voice "Cloning" is Style Transfer

arXiv· arXiv:2605.16578· Kaitlyn Zhou, Federico Bianchi, Martijn Bartelds, Anna Pot, Yongchan Kwon, James Zou·Indexed by this site · 3 days ago· Originally publishedEvaluation
AI summary

A study found that commonly used voice cloning models do not faithfully replicate an individual's voice, but instead apply style transfer to the source voice. Human raters perceived the cloned voices as more authoritative, more likable, more like customer service, and more like a real person than the original voices, and they had greater trust in the cloned voices and were more willing to disclose sensitive personal information to them. The study also showed that voice cloning leads to homogenization of speaker characteristics, with reduced variance in accent, speech rate, and audio embedding space. The authors state that these results reveal new limitations and risks of voice cloning technology and its potential impact on human behavior.

Read originalarxiv.org