When creating unit selection voices for personal use, e.g., for use in communication aids, it is often desirable to keep the speech database as small as possible. The present study examines the effects of database size and database content on the intelligibility of synthetic speech produced by the latest version of the ModelTalker TTS system. Intelligibility here is measured objectively with an open response SU sentence task. While previous work has examined similar questions, that work has typically been with an eye toward completeness of the database coverage and using tasks that assess perceptual quality, but not explicitly intelligibility.
Index Terms: speech synthesis, unit selection, database size, database content, intelligibility, personal synthetic voices
Cite as: Bunnell, H.T. (2010) Crafting small databases for unit selection TTS: effects on intelligibility. Proc. 7th ISCA Workshop on Speech Synthesis (SSW 7), 40-44
@inproceedings{bunnell10_ssw,
author={H. Timothy Bunnell},
title={{Crafting small databases for unit selection TTS: effects on intelligibility}},
year=2010,
booktitle={Proc. 7th ISCA Workshop on Speech Synthesis (SSW 7)},
pages={40--44}
}