The performance of existing speech recognition systems degrades rapidly in the presence of noise. In this paper, we study a solution to this problem using a preprocessor based on the spectral substraction method. We have made a large amount of tests with a recognition system based on hidden Markov models (HMM) approach with special emphasis in a car environment. Results are much better than those obtained without any additional processing. In sorne cases, recogniton rates are as little as when signa/ is not corrupted by noise (error<1 .5%).

Peer Reviewed