Noise compensation for speech recognition using probabilistic models

Abstract

There are significant applications of speech recognition where a poor signal-to-noise ratio seriously degrades performance. In many cases the noise properties are constant or change only slowly. Klatt has proposed a solution based on noise masking and Bridle et al. one on noise marking. Our approach extends their techniques to make fuller use of the available data and we apply our method to recognizers based on whole-word pattern matching using both DTW and Markov model algorithms. The method also gives improved training where it is convenient for this to be done in a noisy environment, or where training in noise is essential because of voice quality changes that are a consequence of the noise.