Stress predicition for Mandarin text-to-speech system using discourse context feature

Abstract

Stress prediction is a vital factor for both speech synthesis and natural speech understanding. In this paper, we investigate how to improve the performance of Mandarin stress predictor by introducing discourse context features. Two widely accepted statistical methods are employed to evaluate given/new status and informativeness of a word which are two major discourse context features in the stress prediction. Syntactic and other sentence context features are used in the method as well. The machine learning algorithm, Maximum Entropy model, is adopted to predict which syllable will be stressed with the textual features. The experimental result shows that the performance of the predictor with the discourse context features is better than the methods using the sentence features and even close to the methods using both the acoustic and textual features.