Development of Story Text-to-speech System Based on Story Genres

Storytelling is a distinct speaking style which embraces various expressive subtleties aimed to draw the attention of the children. Studies have shown that Text-to-speech (TTS) systems have the tendency of not conveying the right emotion expressivity in their speech outputs. The objective of this work is to develop prosody models to capture the story semantics present in the Hindi children stories. In this work, we have considered three prosodic parameters to be modeled: duration, intonation, and intensity of the syllables. We have proposed textual (positional, contextual and phonological) and storysemantic features for modeling the prosody. The proposed prosody models will be integrated with the TTS framework to synthesize speech in storytelling style. We used Classification and Regression Tree (CART), Feed-forward neural network (FFNN), and Support Vector Machine (SVM) for modeling the prosody. We pose the problem of modeling the prosody as a data driven statistical transformation from input text onto the feature space to capture the implicit prosodic and semantic knowledge of the syllables. The evaluation of the models was carried out using objective and subjective measures. A framework is designed to further improve the performance of the prosody models by including the story genre specific prosody modeling. The three genres of the story considered are Folk-tale, Legendary, and Fable. The subjective evaluation of the synthesized story speech shows the proposed framework is capable of improving the storytelling style.