Increasing the speed of unit selection speech synthesis by beam-search and pruning

Concatenation of prerecorded speech segments is one of the speech synthesis methods, which is being named as unit selection based speech synthesis. The basic idea of such systems is to select the best speech units from a large database and concatenating them to produce the final waveform. This process usually is done based on minimizing a predefined cost function. Searching the best data sequence on a vast amount of data is a time-consuming process. Meanwhile, most of the applications cannot stand a long time for production of desired speech. In this paper, the beam search method is introduced to limit the search space and to increase the synthesis speed. In this method, the search tree is pruned to remove the paths which are less probable to be the optimum path. By so doing, the quality of synthesized speech is maintained and the time of synthesis is reduced. Prior to the results of experiments, removing 60% of search paths by beam search causes the time of synthesis to be reduced by 57%, decreasing the naturalness of speech from 4.2 to 4, based on MOS evaluation method. This decrease in the naturalness is equal to only 4.7%. Also by pruning the data based on the two approaches proposed in this paper, synthesis speed is increased by 10% and 14.07% respectively. The naturalness, however, is reduced by 5% and 3.8%.