Although reaching results of almost 100% recognition rate in quiet surrounding, the performance of speech recognizers decreases very fast in a noisy environment. Various algorithms are available for noise reduction and they are working well in reducing moderate and stationary ambient noise. The idea presented here is to continuously adapt templates to different situations and noise scenarios by updating original reference patterns after every successful recognition. A new reference template is calculated from the weighted sum of the recognized reference template and the actual test pattern. In this way, it is possible to continuously incorporate information about different variations of noise and Lombard speech in the reference templates, thus improving recognition performance in changing situations.
Bibliographic reference. Dvorak, Susanne / Hormann, Thomas (1991): "High-performance speech recognition in noise by continuously updated reference templates", In EUROSPEECH-1991, 1375-1378.