Not a Bug, a Feature? How Repetition Enhances Perceived Naturalness in Synthesized Conversation

초록

While recent conversational AI can generate fluent language, it often fails to replicate the interactional naturalness of human discourse. This study investigated the role of a single, ubiquitous feature of human interaction-repetition of a prior speaker's turn-in listeners' perception of conversational naturalness. First, we used conversation analysis to identify and analyze instances of repetitional responses in a corpus of natural human interaction. Second, we conducted a perception experiment where participants listened to audio stimuli in two conditions: one containing the original repetition and a manipulated version with the repetition removed. Our findings show that, even in synthesized speech, conversations containing structural repetition are perceived as significantly more natural than those without it (+1.57on the Likert scale, p < 0.001), suggesting how repetition is not mere redundancy but an important pragmatic and interactional resource that enhances conversational naturalness. While the avoidance of lexical repetition is often a heuristic for improving the quality of generative AI, our results suggest that a more nuanced design philosophy is warranted.

키워드