PHONEME LENGTH DECIDING METHOD IN VOICE SYNTHESIS AND METHOD OF LEARNING PHONEME LENGTH DECISION TREE

摘要

PURPOSE: A phoneme length deciding method in voice synthesis and a method of learning phoneme length decision tree are provided to statistically decide the length of a syllable to generate a natural synthesis sound. CONSTITUTION: Constructive attributes given to each of phonemes that are learning objects are analyzed. The phonemes are classified according to the attributes and an average error of the classified groups is obtained. Among the attributes, an attribute having the smallest average error is determined. The groups classified according to the determined attribute is classified again according to the other attributes, and an average error of the classified groups is obtained. The above-described steps are repeated until the average error becomes smaller than a predetermined value, to acquire a phoneme length decision tree. An average phoneme length of each group is allocated to the end node of the phoneme length decision tree. The length of a phoneme that is an object of voice synthesis is decided with reference to the phoneme lengths.