Reference speech pattern generating method

Claim 1

A reference pattern generating method comprising: a step for inputting a learning speech; a step for extracting spectral parameters of the learning speech in units of frames; a segmentation step for dividing a time series of extracted spectral parameters into segments for each voice interval; a step for clustering the segments into a plurality of clusters; a step for computing, for each cluster, an initial reference pattern representing the cluster; a correction step for correcting the segmentation boundaries of the spectral parameter time series through use of the computed reference patterns; a clustering step for clustering the segments of the spectral parameters corrected in segmentation boundaries into clusters each corresponding to one of the initial reference patterns; and a corrected reference pattern computing step for computing, for each cluster, a reference pattern representing the cluster and repeating the clustering of the learning speech through use of the computed reference patterns until a measure of error is converged, whereby corrected reference patterns are computed; the correction step, the clustering step, and the corrected reference pattern computing step being performed at least once and the reference patterns obtained by the last corrected reference pattern computing step being regarded as reference patterns desired to be obtained.