Unit selection using pitch synchronous cross correlation for Japanese concatenative speech synthesis

We describe a corpus-based approach to improving synthesized speech quality and present two useful cost functions for unit selection. One is pitch-synchronous cross correlation for concatenation costs to reduce the noise caused by phase mismatch at concatenation points. The other is a discontinuous cost function for internal and concatenation costs to eliminate unnecessary cost calculation. An evaluation showed that incorporating pitch-synchronous cross correlation cost was better than using a conventional cost function. In addition, an opinion test to assess the naturalness of the synthesized speech indicated that the proposed method was 0.7 points better on a seven-point MOS(Mean of Opinion Score) than the conventional system. This paper also discusses other improvements in the performance of text-to-speech systems. In this session, we will demonstrate our Japanese text-to-speech system.