Automatic classification of intonational phrase boundaries

Abstract

The relationship between the intonational characteristics of an utterance and other features inferable from its text represents an important source of information both for speech recognition, to constrain the set of allowable hypotheses, and for speech synthesis, to assign intonational features appropriately from text. This work investigates the usefulness of a number of textual features and additional intonational features in predicting the location of one particular intonational feature—intonational phrase boundaries—in natural speech. The corpus for this investigation is 298 utterances from the 774 in the DARPA-collected Air Travel Information Service (ATIS) database. For statistical modeling, we employ classification and regression tree (CART) techniques. We achieve success rates of just over 90%, representing a major improvement over previous attempts at boundary prediction for spontaneous speech.