In this paper, a data-driven approach that compensates the HMM parameters for the noisy speech recognition is proposed. The various statistical information necessary for the HMM parameter compensation is estimated during the HMM training procedure by using the expectation-maximization algorithm. Instead of assuming some statistical approximations for the model combination in the conventional methods such as the PMC, the estimated statistical information is directly combined with the clean speech HMM parameters to produce the compensated parameters. The proposed method has shown improved results compared with the PMC in the isolated Korean word noisy speech recognition experiments.
Cite as: Chung, Y.-J. (2005) A data-driven approach for the model parameter compensation in noisy speech recognition. Proc. Interspeech 2005, 961-964, doi: 10.21437/Interspeech.2005-229
@inproceedings{chung05_interspeech,
author={Yong-Joo Chung},
title={{A data-driven approach for the model parameter compensation in noisy speech recognition}},
year=2005,
booktitle={Proc. Interspeech 2005},
pages={961--964},
doi={10.21437/Interspeech.2005-229}
}