Model generating method, and speech synthesis method and apparatus

Abstract

A method is performed by at least one processor, and includes acquiring training speech data by concatenating speech segments having a lowest target cost among candidate concatenation solutions, and extracting training speech segments of a first annotation type, from the training speech data, the first annotation type being used for annotating that a speech continuity of a respective one of the training speech segments is superior to a preset condition. The method further includes calculating a mean dissimilarity matrix, based on neighboring candidate speech segments corresponding to the training speech segments before concatenation, the mean dissimilarity matrix representing a mean dissimilarity in acoustic features of groups of the neighboring candidate speech segments belonging to a same type of concatenation combination relationship, and generating a concatenation cost model having a target concatenation weight, based on the mean dissimilarity matrix, the concatenation cost model corresponding to the same type of concatenation combination relationship.