High quality text-to-speech conversion requires robust and accurate text processing, well founded modelling of phonological/phonetic processes, and sound generation with high intelligibility and naturalness. The present paper describes a new text-to-speech system under development which provides a flexible framework for integrating these processing modules. We give an overview of the system design and describe some recent advances in the related fields: text processing centered around a large lemma-based lexicon, the rule formalism and data structures of the system, and some recent work on improving voice source modelling in a formant synthesizer.