The Importance of Sub-Utterance Prosody in Predicting Level of Certainty

Abstract

We present an experiment aimed at understanding how to optimally use acoustic and prosodic information to predict a speaker's level of certainty. With a corpus of utterances where we can isolate a single word or phrase that is responsible for the speaker's level of certainty we use different sets of sub-utterance prosodic features to train models for predicting an utterance's perceived level of certainty. Our results suggest that using prosodic features of the word or phrase responsible for the level of certainty and of its surrounding context improves the prediction accuracy without increasing the total number of features when compared to using only features taken from the utterance as a whole.