The evaluation measures for Text-to-Speech (TTS) systems include accuracy and naturalness. While the first measure has received much attention, the second one appears overlooked by most TTS systems. We find that this is mainly due to the inherent limitation of the exclusively stochastic methods in spoken language generation. It is well known that discourse information, which is hard to capture in a stochastic-only system, plays a significant role in reconstructing the naturalness of a given intonation contour. In this thesis, we propose to utilize not only syntax, but also semantics and discourse in implementing a prototype intonation generation system for Korean in a Combinatory Categorial Grammar framework. We test the implemented system by using MBROLA, a diphone-based TTS system.
