Abstract
With development of robotics, research and development for robots which live with human have quickly spread. It is required for these robots to communicate smoothly with users. Therefore, user's emotion distinction from speech, which is one of techniques to estimate user's conditions, has been widely studied. In the conventional methods, they distinguish emotions related with prosodic features or voice quality features which are extracted from words. These features include speaker's personality, and distributions of features differ according to speakers, even though they speak in same emotion. For speaker independent emotion distinction system, individual difference in features causes decrease of distinction rates. Moreover, in the same speaker, several different emotional speeches can have similar features. In this case, distributions of several different emotional speeches overlap each other in the feature space. So, the distinction system can not distinguish emotion of new speech which has been input into the overlapped area.The distinction technique which has robustness against individual difference in features or against overlapped distributions of emotion has been required. This paper proposes new technique to distinguish emotion transition based on change of prosodic features at emotion transition. The mode of prosodic features are assumed to change at the speaker's emotion change. The distinction system distinguishes emotion transition according to an event which is detected at the mode change on prosodic features. Even though emotional speech has been input into overlapped area, the system can distinguish whether emotion transits or not by detecting the events of mode change on prosodic features. In this paper, 4 emotion transitions; “Interest → Sympathy”, “Interest → Joy”, “Interest → Doubt” and “Interest → Antipathy” are studied. In each emotion transition, the event of mode change on prosodic features are extracted, and emotion transition is distinguished. The distinction rates are compared with those of the conventional technique of normalizing features and the usefulness of proposed technique has been confirmed.
