This paper discusses the use of state-based cepstral mean compensation (SBCMC) for transforming a set of HMM word models trained on clean data into a set of models which can be used under a specific set of noise conditions. Two specific methods are described for calculating the cepstral mean corrections: parallel model decomposition (PMD) and Weiner filtering (WF). It is shown that under certain normality assumptions, the WF method is equivalent to the PMD method for the case of zero noise variance. Experimental results are presented using both synthetic and real noise data. In both cases, good performance improvements are obtained from both the PMD and WF methods and there appears to be little to choose between them. The overall conclusion is that SBCMC is a very simple but effective and computationally efficient approach to dealing with noise in a HMM based system.
Cite as: Young, S.J. (1992) Cepstral mean compensation for HMM recognition in noise. Proc. ETRW on Speech Processing in Adverse Conditions, 123-126
@inproceedings{young92_spac,
author={S. J. Young},
title={{Cepstral mean compensation for HMM recognition in noise}},
year=1992,
booktitle={Proc. ETRW on Speech Processing in Adverse Conditions},
pages={123--126}
}