Study of Activation Level Estimation using Phonetic Information during Voice Chat

Abstract

In this paper, we discuss the feasibility of conversational activation level estimation by using phonetic and turn-taking features. At first, we experimentally collected six groups of 3-persons conversational voice at three different activation levels. Then, we calculated the phonetic and turn-taking features, and analyzed the correlation between the features and the activity level. The analysis revealed that response latency, overlap rate and speech rate correlate with activation level and are less sensitive to individual deviation. Then, we formulate multiple regression equation, and examined the estimation accuracy using another six 3-persons groups. The results demonstrated the feasibility to estimate activation level at about 24% root-mean-square-error.