Data-Driven Phrase Break Prediction for Bengali Text-to-Speech System

In this paper, an approach is proposed to accurately predict the locations of phrase breaks in a sentence for a Bengali text-to-speech (TTS) synthesis system. Determining the positions of phrase breaks is one of the most important tasks for generating natural and intelligible speech. In order to approximate the break locations, a feed-forward neural network (FFNN) based approach is proposed in the current study. For acquiring prosodic phrase break knowledge, morphological information along with widely-used positional and structural features are analyzed. The importance of all the features is demonstrated using a model-dependent feature selection approach. Finally the phrase break predicting model is implemented with the selected optimal set of features and incorporated inside a Bengali TTS system built using Festival framework [1]. The proposed FFNN model is developed using the optimally selected morphological, positional and structural features. The performance of the proposed FFNN model is compared with widely used Classification and Regression Tree (CART) model for prediction of breaks and no-breaks. The FFNN model is evaluated objectively on the basis of precision, recall and a harmonized measure - F score. The significance of the phrase break module is further analyzed by conducting subjective listening tests.