Tree-based modeling of prosody for Korean TTS systems

Abstract

To help listeners understand speech, speakers use prosody, which concerns the suprasegmental aspects of spoken language and carries information that is not readily expressed in the literal meaning of the words nor in their syntactic relations. As text-to-speech systems are being incorporated into more and more various applications like e-mail reader and language education system, human users```` desire for a higher quality system is increasing. However, while the current technology makes it possible to obtain a system whose intelligibility is quite high, the lack of natural prosody is the major source of barriers to meeting the users```` expectation. Therefore, it is the one with the greatest need for improvement. The goal in this thesis is to develop a computational model of Korean prosody that improves the naturalness of synthetic speech. We model four prosodic components which are phrasing, loudness, duration, and speech intonation. Then a prosody generation model is incorporated into our Korean text-to-speech system. Our work on prosody modeling can be described by the following theoretical and experimental contribution. First, we suggest a novel Korean prosody structure from an engineering viewpoint to get a more tractable computational model of Korean prosody. Although various theories on Korean prosody structure have been devised, they are rather complicated and difficult to be embodied in a prosody generation module. Thus, we modify the conventional theories and show the appropriateness of the proposed structure in developing a prosody generator. Second, by taking a tree-based framework for prosody modeling, we scientifically discover the linguistic information saliently affecting Korean prosody and draw up rules for the syntax-to-prosody relationship. Since the tree-based framework gives the high comprehensibility in the prediction process, we are able to identify the underlying rules that control prosody by interpreting the trees. Third, we apply boo...