Disfluent speech analysis and synthesis: A preliminary approach

Abstract

Despite of the existence of high quality unit selection speech synthesizers, they are based on a reading style approach. However, new applications such as Speech-to-Speech Translation or Speech User Interfaces demand a talking style which is more natural in these contexts. Disfluencies are a major characteristic of talking style so that it is convenient to be able to generate disfluent speech. In the present paper a preliminary analysis of pitch and segmental duration in repetitions and filled pauses is presented. Simple rules to predict these prosodic features are derived from the previous analysis and used for synthesis. Evaluation shows an increase in naturalness while overall quality is decreased.