Pause structure in synthetic speech

Abstract

Intonation contouring of synthetic speech improves both intelligibility and comprehension [Slowiaczek and Nusbaum, to appear], while flattening appears to interfere with speech perception [Larkey and Danley (1983)]. Given the proven importance of F0, several problems remain in defining the domain of prosodic contour. Among the issues are pauses and intonational domains. These are particularly critical in an unlimited text-to-speech system where input is often unpunctuated long complex sentences. This paper reports on current work to determine pause structure of synthetic speech as a component of specifying the domain of prosodic contouring. The system uses a deterministic bottom-up parser to give a syntactic analysis of a sentence. Based on this and other information, a pause structure is computed algorithmically. The pause-parsed structure then serves as input for later stages in the application of F0. Syntactically based pause insertion is compared with a simple function/content word-based pause-insertion algorithm. We give evidence that the insertion of pauses based on a syntactic parse increases naturalness.