Modification of entropy-based algorithm for determining words boundaries in noisy conditions

Abstract

A modification of the algorithm for finding word boundaries in continuous speech based on an analysis of the entropy of a speech signal is described. At the preprocessing stage, the signal is divided into frames; the entropy value is calculated for each frame and compared with the threshold value of the entropy γ. The algorithm is carried out in two stages, first a rough and then more precise definition of the word boundaries is made. At each stage, its own minimum distance between words is used. A modification is proposed, which consists in the fact that the entropy threshold γ is determined on the basis of the calculated SNR coefficient taking into account the experimental selection of signal parameters at the stage of precise determination of word boundaries in a noisy speech signal. Checking the functionality of the modified algorithm on noisy signals has shown that taking into account the effect of SNR on the selection of signal parameters increases the accuracy of word boundary determination.