Abstract

The computational performance of automatic speech recognition can be enhanced either at the front end by measuring a set of patterns for speech versus silence or noise discrimination, also by speech enhancement or by an efficient classifier at the backend. Employing the speech versus silence or noise discrimination before performing speech enhancement operations plays a vital role, to prevent the over thresholding of silence coefficients to reduce the computational complexity that would not be useful for maintaining the intelligibility of speech. The comparison of experimental results on the database created in the room environment and on the standard TIMIT database reveal that the performance of the proposed system is better under noisy as well as in noise free conditions.