Natural prosody generation for domain specific text-to-speech systems

Abstract

The paper proposes a method for generating natural prosody for text to speech systems. In this method, sentences are composed by inlaying a variable word into each slot in prepared sentence structures. This method can be used for domain specific text to speech applications that don't require so many sentence structures but many words. Important parameters for prosody are declination of F0 contour, accent strength of word, and position and duration of pause. So we construct a database containing these parameters, manually extracted from natural speech samples that have the sentence structures to be synthesized. In the process of prosody generation, these parameters are retrieved by the type of the sentence structure, and the other parameters are generated with rules. The F0 contour is generated by superposing these components on baseline frequency. Mean opinion score of prosody naturalness for the speech synthesized by the proposed method is 3.6. This is 1.2 points better than that of the former method.