In this paper we propose a novel compensation approach named Canonical Correlation Based Compensation(CCBC) to improve the performance of recognizers under noisy environment. In practical speech recognition applications, the mismatching between training and testing environments often seriously diminish recognition accuracy. Because the testing environments are not always known beforehand, The adaptive compensation framework is a practical approach to cope with this problem. While other compensation methods often deal with only some of the cepstrum changes, and is effective only for specific condition, the new approach proposed here is noise independent and can compensate all of the three main differences between two environments, i.e. mean value shift, norm shrink and the bad correlation of each dimension between training and testing speech. The experimental results show that our method has very good compensation effect.
Bibliographic reference. Yu, Dong / Huang, Taiyi (1995): "Canonical correlation based compensation approach for robust speech recognition in noisy environment", In EUROSPEECH-1995, 477-480.