Pronunciation modeling in speech synthesis

This dissertation investigates the area of pronunciation modeling in speech synthesis. By pronunciation modeling, we mean architectures and principles for generating high-quality human-like pronunciations. The term pronunciation modeling has previously been applied in the context of speech recognition (e.g. Byrne et al. 1997). In that context, it describes theories and procedures for handling the pronunciation variation that naturally occurs across speakers. In contrast, our work is in the domain of text-to-speech synthesis, which, as we will show, requires modeling the pronunciation variation of an individual whose speech the synthesizer is attempting to model. We will explain our methodology for learning and reproducing pronunciation variation on an individual basis, and show how most crucial features of such variation can be easily generated using the architecture we describe. Throughout the course of this exposition, we highlight contributions to linguistic theory that such a thorough analysis of individual variation provides. We describe the postlexical module of an English text-to-speech synthesizer. This module is responsible for transforming underlying lexical pronunciations from a lexical database into contextually appropriate surface postlexical pronunciations. This transformation is achieved by machine learning of a corpus of hand-labeled postlexical pronunciations that have been aligned with lexical pronunciations. The machine learning is conducted by a neural network, whose architecture and data encoding we describe. A thorough analysis of the performance of the postlexical module is offered, with attention to the relative success of the neural network at learning a wide range of postlexical phenomena. We examine the extent to which a symbolic approach to allophony is warranted, and provide an acoustic analysis that attempts to provide an answer to this question. Assessments of the success of currently existing theories of phonetics, phonology and their interface are offered, based on the experience of generating a complete postlexical phonology of English for use in synthetic speech. Comments University of Pennsylvania Institute for Research in Cognitive Science Technical Report No. IRCS-98-09. This thesis or dissertation is available at ScholarlyCommons: http://repository.upenn.edu/ircs_reports/55 PRONUNCIATION MODELING IN SPEECH SYNTHESIS