MARY TTS participation in the Blizzard Challenge 2007

Abstract

This paper describes the second participation of the open source MARY TTS unit selection system in a Blizzard challenge. Compared to last year’s system, a number of welldefined changes have been made to the algorithm, concerning unit definition, prosody models, and signal modification. The results in this year’s challenge are considerably improved, confirming that the changes were worthwhile. The paper also reports on an approach to the selection of a subset of the utterances provided, in order to build a voice with good coverage not larger than the pre-defined “Arctic” subset. Results show that this small voice is perceived slightly better than the voice we built from the Arctic subset.