This paper proposes three noise adaptation algorithms which allow improvements in the performance of speech recognition systems under noisy conditions. They are VQ-based feature mapping techniques which hierarchically transform noisy feature vectors into clean feature vectors. The first algorithm was originally an unsupervised speaker adaptation algorithm. It is based on hard clustering and hierarchically adapts the noisy input data to a small set of codebooks created from clean data. The second algorithm is a modified version of the first. It redefines the mapping function using the notion of cluster scope. The last algorithm proposes a fuzzy clustering technique as a substitute to the original hard clustering technique. In the NATO digit task, these algorithms significantly improve the performance of CRIM's speech recognition system.
Cite as: Cung, H.M., Normandin, Y. (1992) Noise adaptation algorithms for robust speech recognition. Proc. ETRW on Speech Processing in Adverse Conditions, 171-174
@inproceedings{cung92_spac,
author={H. M. Cung and Y. Normandin},
title={{Noise adaptation algorithms for robust speech recognition}},
year=1992,
booktitle={Proc. ETRW on Speech Processing in Adverse Conditions},
pages={171--174}
}