A complete Croatian language text-to-speech system

This paper presents the first attempt at building a complete diphone based text-to-speech system for the Croatian language using the Festival environment. To collect the target diphones we generated 855 nonsense words and recorded their pronunciations. Diphones were labeled both automatically and manually. Modules for textual and linguistic analysis were developed and discussed. A speech quality evaluation proved high intelligibility − 87.5% of all synthesized words were understood correctly by testers.