Feasibility of pervasive monitoring of nonverbal information in daily office activity

To allow robots and agents to serve users with less-distractive timing, the systems need to estimate the communication status of human social activity. Nonverbal information has the potential to be used to estimate communication status without recording conversational content. In this study, we prototyped a pervasive speaker estimation system and examined the feasibility of automatic estimation of nonverbal information. Analysis of 5-hour of unconstrained activity in a laboratory demonstrated 90% speaker ID estimation accuracy and a less than 8.5% error rate for the durations and frequencies of utterances excluding short utterances of less than 1 s. The feasibility of automatic analysis of turn-taking patterns in daily office activity has also been suggested.