Synthesizing Suprasegmental Speech Information Using Hybrid of GA-ACO and Dynamic Neural Network

In generating natural speech by machines, removing the suprasegmental information (such as stress, timing and pitch frequency) results in unpleasant speech. To provide this information for synthesizing natural speech in Farsi language, a dynamic neural network (DNN) is used in this study. The inputs of DNN are word-level and syllable-level features as part of speech tags, word length, and type of punctuation mark at the word-level, and type of vowel and consonants, and position indicator of syllable at the syllable-level. To reduce the number of inputs of DNN, hybrid of genetic algorithm (GA) and ant colony optimization (ACO) is used for feature selection. The output layer of DNN includes nine nodes which provide suprasegmental information at the syllable level including pitch contour, log-energy level, duration information and pause data. Simulation results show that suprasegmental information is predicted with low root mean square error by using this hybrid soft-computing model.