In spoken dialogue systems, especially chatting systems, smooth turn-taking function is one of the most important techniques to realize natural interaction with users. Fillers are presumed to play an important role in natural turn-taking in human dialogue. To this end, in this study, we analyzed a dialogue corpus where people talk with a humanoid android ERICA which was remotely operated. At first, we examined the occurrence rate of fillers at clause units in the case of turn-taking or not. The results suggest that we can see some differences among the roles of the participants. Next, some functions of each filler form were analyzed, which suggests a possibility that forms of fillers can be predicted using an adjacency pair.
