Review of noise robust automatic speech recognition

This report presents a review of the main research directions in noise robust automatic speech recognition. The widespread use of (mobile) applications in time-varying noisy conditions has promoted robust ASR to a major research issue. The inclusion of some form of compensation against any kind of noise interference in ASR is indispensable to guarantee an acceptable level of recognition accuracy in these situations. First, the basic structure of a typical ASR system is recalled, followed by a definition of the noise robustness problem. We then proceed with an extensive overview of the most important solutions to this problem. Three main categories are distinguished : (1) speech enhancement, (2) robust feature extraction, and (3) model adaptation techniques. Finally, we briefly discuss some related issues like robustness against variability in pronunciation and speaking style.Hermus K., Wambacq P., ''Review of noise robust automatic speech recognition'', Internal report PSI-SPCH-05-1, K.U.Leuven, ESAT, January 2005, Leuven, Belgium.nrpages: 21status: publishe