Abstract
This paper addresses the issue of automatic word/sentence boundary detection in both quiet and noisy environments. We propose to use an entropy based contrast function between the speech segments and the background noise. A simplified data based scheme of computing the entropy of the speech data is presented. The entropy-based contrast exhibits better-behaved characteristics as compared to the energy-based methods. An adaptive threshold is used to determine the candidate speech segments, which are subjected to word/sentence constraints. Experimental results show that this algorithm outperforms energy-based algorithms. The improved detection accuracy of speech segments results in at least 25% improvement of recognition performance for isolated speech and more than 16% for connected speech. For continuous speech, a preprocessing stage comprising of the proposed speech segment detection makes the overall HMM based scheme more computationally efficient by rejection of silence periods.