Due to the inconsistency between HMM training and synthesis application in HMM-based speech synthesis, the minimum generation error (MGE) criterion had been proposed for HMM training. This paper continues to apply the MGE criterion for tree-based clustering of context dependent HMMs. As directly applying the MGE criterion results in an unacceptable computational cost, the parameter updating rules of the MGE criterion are simplified to rapidly update the parameters of clustered models, and an appropriate strategy by combining the MGE criterion with the ML criterion is designed to select the optimal question for tree node splitting. From the experiment results, the quality of synthetic speech was improved after applying the MGE criterion for HMM clustering.
Cite as: Wu, Y.-J., Guo, W., Wang, R.-H. (2006) Minimum generation error criterion for tree-based clustering of context dependent HMMs. Proc. Interspeech 2006, paper 1373-Wed3BuP.6, doi: 10.21437/Interspeech.2006-401
@inproceedings{wu06c_interspeech,
author={Yi-Jian Wu and Wu Guo and Ren-Hua Wang},
title={{Minimum generation error criterion for tree-based clustering of context dependent HMMs}},
year=2006,
booktitle={Proc. Interspeech 2006},
pages={paper 1373-Wed3BuP.6},
doi={10.21437/Interspeech.2006-401}
}