Automatically Derived Discourse Segmentation Algorithms Based on Acoustic-Prosodic Features

We describe an investigation aimed at furthering the understanding of how speakers communicate discourse structural information using intonational features. We used the read and spontaneous speech of two speakers from the Boston Directions Corpus (BDC) to automatically identify elements of discourse structure based on intonational features. Unlike previous acoustic-prosodic analyses of discourse corpora, we used discourse segmentations produced by both naïve and expert labelers and analyzed more than a single speaker. We extended previous work by using machine learning algorithms to automate the construction of models that identify discourse segment initial (SBEG), final (SF), and medial (SCONT) phrases. Though the models used a variety of features for classifying all three elements, general trends exist. The length of the preceding pause and the change in the maximum fundamental frequency (f0) from the previous phrase were the most informative features for classifying SBEG phrases. For SF phrases, important features included the duration of the subsequent pause and f0 average. Finally, SCONT phrases were more difficult to classify automatically. Their classification often involved constraints on the duration of the preceding and subsequent pauses and the change in f0 maximum from the previous phrase. The learned models provide further insight into how speakers use acoustic-prosodic properties to convey information about the structure of discourse. A firmer understanding of the relationship between acoustic-prosodic cues and discourse structure would lead to improvements in systems for interpreting and generating dialogues and in applications such as text-to-speech synthesis.