In this paper, task-specific adaptation is proposed to improve Chinese name recognition performance. Since acoustic models are usually trained using large vocabulary continuous speech corpora, there exists distortion between modeling and decoding in name recognition. To compensate the mismatch, task-specific adaptation, which is performed in the MLLR framework with multi-regression classes, is proposed. Experimental results show that task-specific adaptation is very effective in Chinese name recognition to compensate the mismatch.
Cite as: Ding, G., Xu, B., Wang, X., Cao, Y., Ding, F., Tang, Y. (2004) Task-Specific Adaptation in Chinese Name Recognition. Proc. International Symposium on Chinese Spoken Language Processing, 261-264
@inproceedings{ding04b_iscslp,
author={GuoHong Ding and Bo Xu and Xia Wang and Yang Cao and Feng Ding and Yuezhong Tang},
title={{Task-Specific Adaptation in Chinese Name Recognition}},
year=2004,
booktitle={Proc. International Symposium on Chinese Spoken Language Processing},
pages={261--264}
}