Mandarin stress detection using syllable-based acoustic and syntactic features

Automatic stress detection is important for both speech understanding and natural speech synthesis. In this paper, we report on experiments with several classifiers trained on a hand-labeled corpus, using acoustic, lexical and syntactic features. Results show that boosting neural network (NN) classifier achieves the best performance for modeling acoustic features, and that conditional random fields (CRFs) is more effective for lexical and syntactic features. The combination of the acoustic and syntactic classifiers yield 84.23% stress detection accuracy rate. When comparing with previous work on the same training set and test set, our proposed models have better performance.