TAKING INTO ACCOUNT THE MULTIFACTORIAL CHARACTER OF VOICE CHARACTERISTICS IN THE PROBLEMS OF SPEAKER IDENTIFICATION

When testing the most advanced speaker identification systems on specialized databases, their minimum efficiency, estimated by the error probability at the point of intersection of the error curves, is only a few percent. However, many factors are known that affect the variability of the characteristics of the speaker's voice, each of which has its own, different from the others, influence on the results of the speaker's identification by the characteristics of the voice. The complexity of creating and testing speaker identification systems is the need to quantitatively formalize a number of specific factors that affect the characteristics of his voice. The article discusses the proposed method for accounting for a variety of factors affecting the parameters of the characteristics of the speaker's voice, which provides the fundamental possibility of indirectly accounting for their practically unlimited number. According to this method, «atomic» structures are distinguished from speech signals, which depend on the totality of the main factors that affect the speaker's identification process. With this method, all significant factors affecting the characteristics of the voice will be indirectly taken into account at the level of these structures. Subsequent decisions are made on the combinatorial set of a huge number of these «atomic» structures. «Atomic» speech structures are understood as the spectra of any fragments of any vowel sounds allocated in a time window of 20 ms. «Atomic» structures are selected automatically. The proposed method provides a rational consideration of the multifactorial influence of various parameters, since the spectra of these structures are influenced by all the main factors that characterize the individuality of the voice of a particular speaker. The decision on the identity of the voices of the announcers recorded on different phonograms is carried out on the basis of combinatorics of «atomic» spectra of vowel sounds in both phonograms. The method has shown high efficiency in the examination of phonograms of short duration.