Segmental reliability weighting for robust recognition of partly corrupted speech

Abstract

A new method for the recognition of speech that is partly corrupted by noise is proposed. For this purpose, the reliability of each speech segment is measured based on a log likelihood ratio. A segment that has higher reliability is given more importance in the decision making by a modified Viterbi algorithm. Experimental results showed that the method significantly improves the performance of isolated word recognition under various burst noise situations.