Recognition of Paralinguistic Information using Prosodic Features Related to Intonation and Voice Quality

Besides the linguistic (verbal) information conveyed by speech, the paralinguistic (nonverbal) information, such as intenning the classification of paralinguistic information. Among the several paralinguistic items extions, attitudes and emotions expressed by the speaker, also convey important meanings in communication. Therefore, to realize a smooth communication between humans and spoken dialogue systems (such as robots), it becomes important to consider both linguistic and paralinguistic information. There is a lot of past research concerpressing intentions, attitudes and emotions, most previous research has focused on the classification of the basic emotions, such as anger, happiness and sadness (e.g., Fernandez et al., 2005; Schuller et al., 2005; Nwe et al., 2003; Neiberg et al., 2006). Other works deal with the identification of attitudes and intentions of the speaker. For example, Fujie et al. (2003) report about the identification of positive/negative attitudes of the speaker, while Maekawa (2000) reports about the classification of paralinguistic items like admiration, suspicion, disappointment and indifference. In Hayashi (1999), paralinguistic items like affirmation, asking again, doubt and hesitation were also considered. In the present work, aiming at smooth communication in dialogue between humans and spoken dialogue systems, we consider a variety of paralinguistic information, including intentions, attitudes and emotions, rather than limiting our focus to the basic emotions. The understanding of paralinguistic information becomes as important as linguistic information in spoken dialogue systems, especially in interjections such as “eh”, “ah”, and “un”. Such interjections are frequently used to express a reaction to the conversation partner in a dialogue scenario in Japanese, conveying some information about the speaker’s intention, attitude, or emotion. As there is little phonetic information represented by such interjections, most of the paralinguistic information is thought to be conveyed by its speaking style, which can be described by variations in prosodic features, including voice quality features. So far, most previous research dealing with paralinguistic information extraction has focused only on intonation-related prosodic features, using fundamental frequency (F0), power and duration (e.g., Fujie et al., 2003; Hayashi, 1999). Others also consider segmental features like cepstral coefficients (e.g., Schuller et al., 2005; Nwe et al., 2003). However, O pe n A cc es s D at ab as e w w w .in te ch w eb .o rg