Generation of suprasegmental information for speech using a recurrent neural network and binary gravitational search algorithm for feature selection

Abstract

Suprasegmental (prosody) features of discourse provide a vehicle by which speakers reflect their mental purposes to listeners. Generating suitable prosody information is critical to expressing messages and improving the intelligibility and naturalness of synthetic speech. Generic prosody generators should provide information about pitch frequency (F 0) contours, energy levels, word durations, and inter-word pause durations for speech synthesizers. The present study used a recurrent neural network (RNN) for prosody generation. The inputs of this RNN were word-level and syllable-level linguistic features. To provide data efficiently for the RNN-based prosody generator in the training, validation, and test phases, automatic segmentation and labeling of phonemes were performed. The number of inputs to the RNN was reduced by employing a binary gravitational search algorithm (BGSA) for feature selection (FS). The proposed prosody generator provided 12 output prosodic parameters for the current syllable for representing pitch contour, log-energy contour, inter-syllable pause duration, duration of syllable, duration of the vowel in the syllable, and vowel onset time. Experimental results demonstrated the success of the RNN-based prosody generator in synthesizing the six prosodic elements with acceptable root mean square error (RMSE). By using a BGSA-based FS unit, a lighter neural model was achieved with a 53 % reduction in the number of weight connections, producing RMSEs with acceptable degradation over the no-FS unit prosody generator. The performance of the BGSA-based FS method was compared with a binary particle swarm optimization (BPSO) algorithm, and the BGSA showed slightly better results. A modified mean opinion score scale was used to evaluate the intelligibility and naturalness of synthesized speech using the proposed method.