This study examined how native Korean listeners perceive emotional speech produced by professional human voice actors versus speech generated by Zyphra’s Zonos-v0.1, an advanced AI TTS system. 87 Korean participants listened to 96 audio stimuli, derived from 12 Korean sentences designed to evoke six core emotions and neutral expressions, each rendered by both human actors and AI voices cloned from those actors. Participants assessed each sample for voice identity (human or AI), naturalness, pronunciation, intonation, perceived emotion, and emotional appropriateness. Results revealed that while listeners accurately identified AI voices as artificial (78.81%), they often misclassified human voices as AI (53.90%), especially for neutral speech, indicating a narrowing perceptual gap. Voices perceived as human received higher ratings for naturalness, pronunciation, and intonation, regardless of their true origin. AI voices were rated lower, mainly due to less convincing emotional expression and intonation, though pronunciation was rarely an issue. Human voices were also much more likely to have their intended emotion correctly identified (65.54% vs. 31.66% for AI), and their emotional appropriateness was rated higher. In conclusion, while AI voices have advanced to be confused with human speech, limitations persist in achieving emotional authenticity and contextual appropriateness.
Ryu et al. (Wed,) studied this question.