A Neural Learning Approach for Duration Parameter Generation in Mandarin Speech Synthesis

In this paper, a neural learning approach is investigated, which is designed to generate duration parameter for mandarin speech synthesis. Unlike traditionally used rule-based methods, the novelty of this method lies in that it combines neural learning strategy and prior linguistic knowledge to obtain duration parameter. Rules generalized by linguists are used to encode input vectors of the neural network, and five multi-layer neural networks are built to determine the duration parameter for each tonal syllable. Experiment results show that it is a flexible and effective way to determine duration parameter, and perhaps it provides a helpful way of thought to obtain other prosodic parameters for speech synthesis.