Mansour Sheikhan

To reduce the number of input features to a prosody generator in natural speech synthesis application, a hybrid of an evolutionary algorithm and a swarm intelligence-based algorithm is used for feature selection (FS) in this study. The input features to FS unit are word-level and syllable-level linguistic features. The word-level features include punctuation information, part-of-speech tags, semantic indicators, and length of the words. The syllable-level features include the phonemic structure and position indicator of the current syllable in a word. A modified Elman-type dynamic neural network (DNN) is used for prosody generation in this study. The output layer of this DNN provides prosody information at the syllable-level including pitch contour, log-energy level, duration information, and pause data. Simulation results show that the prosody information is predicted with an acceptable error by this hybrid soft-computing method as compared to Elman-type neural network prosody generator and binary gravitational search algorithm-based FS unit.

speech synthesis, genetic algorithm, ant colony optimization, neural network, prosody