A hybrid statistical/RNN approach to prosody synthesis for Taiwanese TTS

In this paper a hybrid approach which incorporates statstcal modeling of prosodic parameters into recurrent neural network (RNN)based prosody synthesis for Min-Nan speech (Taiwanese) is proposed. It takes syllable as the basic synthesis unit and constucts statistcal models for syllable initial duraton, syllable final duraton, intersylable pause duration, pitch contour of syllable, and log-energy level of syllable. In the training, it normalizes prosodic parameters by these statistical models and uses the results to train an RNN prosody synthesizer. In syntiesis, it denormalizes the RNN outputs by the same statstical models to generate all prosodic parameters requi red by the TTS syst em. Tbadvant age of the appoach can be justified as to relieve the RNN prosody synthesizer of some affecting factors via taking care them by using the statistical models.