Studying the feature space for the description of segments of speech signals in speech recognition problems

Abstract

The article considers a set of features for the description of segments of speech signals with regard to their informative value. Feature sets are used in recognition problems. Effectively solving recognition problems largely depends on the information content of the feature set. The information content of a set of features is estimated using information entropy. In this article, we compare information entropy of the feature sets obtained using the Fourier method and the sub-band method. The article also considers modifications to these methods, namely the transition to the mel scale, used to analyse speech signals. In addition to the mel scale, other frequency axis distortions are considered for the sub-band method. The article presents the results of calculating the information entropy performed using different methods of obtaining a set of features describing segments of speech signals. The study demonstrated that the information content of the feature set increases when using the sub-band method with a nonlinear division of the frequency axis into analysed bands.