摘要 |
PROBLEM TO BE SOLVED: To provide a technique of accent phrase boundary estimation capable of precisely estimating an accent phrase boundary on large speaker-dependent voice data with no phrase boundary label using small speaker-dependent voice data with a phrase boundary label.SOLUTION: The likelihood of a case where word boundary is an accent phrase boundary and the likelihood where no word boundary is included are calculated by inputting an F0 characteristic in the word boundary in a speaker-dependent voice data with no phrase boundary label using a model of speaker dependent phrase boundary F0; a likelihood difference therebetween is calculated to establish a speaker-dependent word nature database using a word nature corresponding to a word boundary the absolute value of the likelihood difference of which is larger than a threshold value and a temporary phrase boundary label. A phrase boundary word model generation section generates a speaker-dependent phrase boundary word model which models the relation between a temporary phrase boundary label and the word nature using speaker-dependent word nature database as a learning data. |