Verification of the factors of individuality through avatar's speech generation system

In recent years, extensive research has been conducted on avatars, and multiple studies have demonstrated their effectiveness as a medium for remote operation. While avatars are effective when teleoperated, they must also be capable of autonomous behavior in the absence of an operator. In particular, avatars whose appearance closely resembles that of a real individual need to possess conversational abilities that reflect the personality of the person being modeled. This paper presents the development of a speech generation system that produces personality-consistent utterances using a large language model (LLM) and speech synthesis technology. We call this system AvatarLLM. Through system evaluation, we examined the factors contributing to the perception of individuality. Experimental results indicated that the utterances generated by AvatarLLM were perceived as more likely reproducing the modeled individual than those of the actual person. Furthermore, we found that the perceived identity of the utterances could influence the perceived identity of the voice itself.