Abstract
In this paper, we study several kinds of nonverbal voices such as cries, screams and laughs. Nonverbal voices are the most primitive and typical emotional voices, and have much Kansei information. They are important from the viewpoint of communication between speakers and listeners. Especially, they are very important tools for communication for babies and children whose faculties of speech are not developed sufficiently. Therefore, it is a very significant subject to analyze nonverbal voices in order to investigate how Kansei information is expressed in them and how listeners recognize speakers' emotions through them. Recently, techniques for analyses of voices from the viewpoint of nonlinear deterministic dynamics have become a center of attraction. Some works suggest that the “nonlinearities”are strongly related to information in voices except linguistic information, especially Kansei information. It should be a challenging problem to tackle this suggestion. In order to approach this suggestion, it might be an important problem to extract Kansei information in nonverbal voices by using the nonlinear characteristics. In this paper, as a first step toward such a nonlinear analysis of nonverbal voices, we carry out a statistical testing to confirm the existence of nonlinearities in voices of cries, screams and laughs using the method of surrogate data. The largest Lyapunov exponent and prediction error by the locally linear method are used as discriminating statistics for testing. As a result, we confirmed the existence of nonlinearities in voices of cries and laughs statistically. Therefore, it was suggested that not only linear but also nonlinear methods are required for extracting much information from at least these two kinds of nonverbal voices. We also discuss functional interpretations of nonlinearities in nonverbal voices.
