Topic transitions and durational prosody in reading aloud: production and modeling

Abstract

The linguistic structure of an utterance is known to affect the durational prosody of sounds, words and phrases. There has been increasing interest in how discourse-level organization affects prosody, in part because modeling discourse-level effects could improve the comprehensibility of longer passages of synthesized text. The approach taken here is to look at how topics are sequenced in a text, and how this affects durational prosody when that text is read aloud. Two speakers of American English were recorded reading a set of text materials on 10 separate occasions. Measurements of these recordings indicated that the type of transition in topic between two successive sentences had a significant effect on the amount of sentence-final lengthening, the duration of the pause between sentences, and the speech rate at the end of a sentence and the beginning of the following sentence. These measurements were then used to create a mathematical model of one speaker, and to generate several versions of one of this speaker's original recordings, with each version incorporating different manipulations of the durational patterns and their variability. These versions were played to listeners, who preferred those where the manipulations included durational patterns reflecting the organization of topics in the text.