Subband Identification of Word Fragments of Speech Signal Word Segments Precedent

Abstract

The article deals with the problem of extracting in the records of speech signals fragments generated during the pronunciation of a certain word form, which is of interest from the standpoint of the applied problem of analyzing the content of speech information exchange. In this case, it is assumed that initially the class of the desired word form is given by the fragment present in the actual recording of the speech signal. Therefore, the problem under consideration, naturally, is called precedent identification. The relevance of developing methods and algorithms for automatic precedent identification of fragments of speech signal recordings in such a formulation is determined by the breadth of their possible applications, for example, in information and analytical security systems. The main factors that determine the complexity of solving this problem are the variability of the properties of fragments of speech signals, even generated when the same word form is pronounced by the same person, and the need to learn from one precedent when determining critical areas of decisive functions. In this paper, it is shown that subband analysis is an adequate basis for solving the problem under consideration and an original mathematical apparatus for its implementation has been developed. Based on the original subband representations, decision procedures for identifying fragments of speech signal recordings, including the selection of pauses between word fragments, have been developed. In particular, training procedures are proposed for one precedent with the preservation of its original subband properties.

Acknowledgments
The research was supported by RFBR grant No. 20-07-00215 a.