Assessing the performance of three methods for separating non-spontaneous and spontaneous speech through simulation

The ability to distinguish spontaneous from non-spontaneous speech can prov e helpful, such as in forensic evidence situations, sorting voice-mail resp onses from voice-mail menus, and automatic segmentation of spontaneous resp onses from prepared questions. The latter situation occurs when trying to c reate a database of spontaneous data from data of a speaker responding spon taneously to prepared prompts. This paper outlines and compares three metho ds for automatically classifying spontaneous and non-spontaneous speech, an d presents the experimental results of the performance of all three methods , evaluated on high quality simulated data. All three methods are based on an analysis of the probability distributions of prosodic features extracted from speech signals.