Detail of publication

Abstract

In this paper, first experiments on statistical parametric HMM-based speech synthesis for the Czech language are described. For speech representation, two different analysis/synthesis methods were employed: traditional Mel cepstral analysis and a high-quality analysis/synthesis method STRAIGHT. Regarding the prosodic and linguistic characteristics of the Czech language, a basic set of contextual factors was proposed. Our experiments showed that disregarding syllabic structure of speech has an insignificant influence on resulting speech quality. The effect of training data amount was also studied. Results indicate, that our experimental HMM-based TTS system can produce speech of a similar quality as unit selection-based TTS system trained with a larger amount of speech data. Furthermore some simple experiments with the adaptation of trained HMMs were performed. In this manner new voices could be obtained with significantly lower amount of speech data.