Constructing scalable TTS system based on Corpus approach

Pruning redundant synthesis instances or tailoring TTS voice font is an important issue of Corpus-based TTS. But pruning redundant synthesis instances, usually results in loss of non-uniform. In order to solve this problem, this paper proposes the concept of virtual non-uniform. According to this concept and the synthesis frequency of each instance, the algorithm named StaRp-VPA is constructed as to make up the loss of non-uniform. In experiments, the naturalness scored by MOS remains almost unchanged when less than 50% instances are pruned off, and the MOS does not severely degrade when reduction rate is above 50%.