Transformation of F0 contours for lexical tones in concatenative speech synthesis of tonal languages

Concatenative speech synthesis (CSS) provides the greatest naturalness. However, it requires a huge stored database resulting a huge footprint. Reducing the capacity of stored database while preserving the quality of CSS, or improving the quality to size ratio (QSr), is still a challenge. In this paper, we propose a method of transforming fundamental frequency (F0) contours of lexical tones, developed from TD-GMM framework that successfully applied for transforming spectral sequence in previous researches, in order to improve the QSr of CSS of tonal languages that results CSS available with limited data at offline stage, storing small online footprint, while preserving perceptual quality. The experimental results show that the proposed F0 transformation outperforms conventional and state-of-the-art F0 contour transformations for transforming lexical tones in terms of speech quality. When applying the proposed F0 contour transformation for transforming lexical tones in CSS of tonal languages, the QSr is enhanced compared with the method of simple F0 exchange while the quality of synthetic speech is preserved.