In this paper, a vocabulary trim down algorithm is proposed in decision tree-based acoustic model to make the model more close to the given task. Using this trim down model as seed model to do task adaptation is also presented. Based on this framework, users can configure the acoustic model by themselves according to their resources (such as vocabulary knowledge, a little amount task specific data, the model size, etc.). Experimental results show that the vocabulary trim down algorithm made the model size being cut off 70% with almost the same accuracy of general model. After adapted by 143 minutes task specific data 27% word error rate reduction can be achieved comparing with the retrained model (using original general purpose data plus all available task specific data) in our Farewell99 dialog system.
Cite as: Guo, Q., Yan, Y., Yuan, B., Zhang, X., Jia, Y., Liu, X. (2000) Vocabulary-based acoustic model trim down and task adaptation. Proc. 6th International Conference on Spoken Language Processing (ICSLP 2000), vol. 4, 109-112, doi: 10.21437/ICSLP.2000-763
@inproceedings{guo00b_icslp,
author={Qing Guo and Yonghong Yan and Baosheng Yuan and Xiangdong Zhang and Ying Jia and Xiaoxing Liu},
title={{Vocabulary-based acoustic model trim down and task adaptation}},
year=2000,
booktitle={Proc. 6th International Conference on Spoken Language Processing (ICSLP 2000)},
pages={vol. 4, 109-112},
doi={10.21437/ICSLP.2000-763}
}