In this paper we propose a novel phrase break prediction model for Mandarin speech synthesis. It is generalized linear models (GLM) with stepwise regression solution. We assume phrase break obeys Bernoulli distribution and then model phrase break probability by Logistic GLM. The attribute set is automatically selected by stepwise regression, which is a totally data-driven method. We also introduce speaking rate as a new attribute for prediction. The proposed method is applied to 2,150 utterances of the Mandarin speech corpus, and it achieves 5.4% higher performances than CART method in open test. The method can be extended to include more linguistic and prosodic attributes and it is very compact for application.
Cite as: Yi, L., Li, J., Lou, X., Hao, J. (2006) Phrase break prediction using logistic generalized linear model. Proc. Interspeech 2006, paper 1468-Tue3BuP.4, doi: 10.21437/Interspeech.2006-384
@inproceedings{yi06b_interspeech,
author={Lifu Yi and Jian Li and Xiaoyan Lou and Jie Hao},
title={{Phrase break prediction using logistic generalized linear model}},
year=2006,
booktitle={Proc. Interspeech 2006},
pages={paper 1468-Tue3BuP.4},
doi={10.21437/Interspeech.2006-384}
}