A comparison of two prosody modelling approaches for Sesotho and Serbian

Abstract

© Springer International Publishing Switzerland 2014. Accurate prediction of prosodic features is one of the critical tasks within a text-to-speech system, especially for under-resourced languages with complex lexical prosody. For synthesized speech to have a natural-sounding in-tonational contour, an adequate prosodic model should be employed. This study compares the Fujisaki model and the HMM-based prosodic modeling in the context of text-to-speech synthesis, for two quite distant languages with rich prosodic systems: Sesotho, a tonal language from the Bantu family, and Serbian, a South-Slavic language with pitch accent. The results of our experiments suggest that, for both languages, the Fujisaki model outperforms the HMM-based model in the modelling of the intonation contours of utterances of human speech.