Pitch behavior detection for automatic prominence recognition

Abstract

In this paper a non-supervised approach for automatic syllable prominence recognition is presented. Previous research in this field showed that syllable nuclei energy and duration are the main cues for prominence detection. The role of the fundamental frequency has also been investigated in the past but was considered secondary or irrelevant for this task. The proposed system uses the energy and the duration of the nucleus while taking into account also the pitch behavior. The algorithm was tested by comparing its results with the annotations of two human experts and a 5.6% accuracy increase with respect to the system not using the pitch behavior was found.