To prevent a synthetic voice from being flat and unnatural by generating a voice synthesis dictionary in which pitch variation is emphasized.

A pitch data string extraction part 15 extracts a pitch data string from voice data extracted from a voice database 23. A pitch edition part 17 edits pitch data so that variation in the pitch data is more remarkable. A phoneme HMM learning part 19 uses the pitch data edited by the pitch edition part 17 for learning and finds out correspondence relation between a phoneme label and phoneme pitch data. A data writing part 21 writes the correspondence relation between the phoneme label and the phoneme pitch data after learning in the voice synthesis dictionary 25.

COPYRIGHT: (C)2008,JPO&INPIT