Speech segment network approach for optimization of synthesis unit set

In this paper, a speech segment network approach for the construction of a suitable synthesis unit set with which high-quality speech can be synthesized, and yet which is of small enough size to be practical, is proposed. The speech segment network approach selects a synthesis unit set in which segmental and/or inter-segmental distortions are minimized by using combinatorial optimization methods such as iterative improvement and simulated annealing. Experimental results using diphone segments have shown that the suitable diphone unit sets, with total or maximum of inter-segmental distortion reduced by about 35 and 30%, respectively, can be constructed using this method. This reduction rate was enhanced as the segment candidate population increased. Effectiveness of this unit set design was also perceptually confirmed by a listening test, using speech synthesized with the selected diphone unit set.