The mood in a conversation is estimated by examining the speakers' affective states. Developing the automatic mood recognition system is one of the bigest issues in smooth communication research of human-human and human-computer/robot interactions. In this research, the affective rating data in UUDB (Utsunomiya University Spoken Dialogue Database) is used as the speakers' affective states. Twenty participants heard speech data and they were asked to rate the mood for each five utterances of the dialogue in the UUDB. The head-counts of participant, who considered the mood as bad/negative for the utterance blocks, are modeled by the Poisson regression modeling technique. The results demonstrate that the model is able to estimate the mood in a conversation by utilizing the speakers' affective states before the target utterance block. Thus the current study indicates that it is possible for a computer/robot to infer mood and it is possible to have effective human-to-computer/robot communications.