Speech Signal Segmentation into Vocalized and Unvocalized Segments on the Basis of Simultaneous Masking

This paper touches upon a model of simultaneous acoustic masking, which detects speech signal components perceived by a human's auditory system. A simultaneous masking algorithm on the basis of this model is proposed. It is shown that, after simultaneous masking, a signal becomes a binary structure that reflects the harmonic structure of a vocalized sequence. It is experimentally proven that this structure can be used to detect key speech segments (from the standpoint of perception by an auditory system). This structure serves as a basis for an algorithm of high-quality segmentation of a speech signal into vocalized and unvocalized segments, which does not require learning before use. The joint use of the algorithms for simultaneous masking and speech signal segmentation is tested, and their performance is evaluated.