A Novel Approach to Speech Signal Segmentation Based on Time-Frequency Analysis

Abstract

The accuracy of speech signal segmentation depends directly on the parameters used to determine the boundaries of the beginning and the end of informative fragments in a continuous speech stream. The purpose of the work is to increase the efficiency of speech/pause segmentation due to the frequency-time analysis of speech signals. A novel original approach to speech/pause segmentation based on the analysis of the values of the mean frequency (in the frequency domain) and short-term energy of the Teager operator function (in the time domain) is proposed. The proposed approach is unique due to an auxiliary algorithm to correct speech/pause segmentation errors, developed on the basis of physiological functioning of the respiratory apparatus organs during the formation of a continuous speech stream. A brief overview of speech signal informative parameters used for speech/pause segmentation has been presented, and the proposed approach performance has been detailed. The suggested approach has been compared with the known methods of speech/pause segmentation for pure and noisy speech signals. The research findings have evidenced the best results of speech/pause segmentation for pure and noisy speech signals being achieved by the methods based on the proposed approach; the ratio of the short-term energy of the Teager operator function to the mean frequency as an informative parameter ensuring maximum relevance to the segmentation problem; an auxiliary algorithm to correct false states enhancing the efficiency of segmentation.