Prosodic Modeling in Text-to-Speech Synthesis

Abstract

This paper discussesthree broad obstaclesthat must be overcome to improve prosodic quality in text-to-speech systems. First, direct and indirect limits set by the signal processing ("synthesis") components. Second, combinatorial and statistical constraints inherent in generalizing from training corpora to unrestricted domains, and that require the integration of content-specific knowledge and detailed mathematical modeling. Third, the nature of many empirical research issues that must be solved for prosodic modeling to improve: they are often too focused and modeldependent for academe, and too long-term for development organizations. 1.