Work

A series of works is devoted to the research of speaker identification characteristics, automated methods for their calculation and the actual task of control system of weakly-formalized object. These weakly-formalized objects are speech and texts created by man. The work based on audio information, but developed complex models are able to maintain information from different sources and of kinds. Developed in the framework of approaches, methods and algorithms for software modules have been used in experimental text-independent speaker identification system. The main otherness of text-independent identification is that it does not require the pronouncing of certain beforehand given phrase or word, because the text-independent system segments the speech fragment by itself, classify segments and builds a comprehensive model of the speaker in these segments, given their connection to a particular class. The authors proposed the division of space based on the acoustic characteristics of phonetic features of speech sounds. The diphone speech recognition method was applied to enable processing of sounds belonging to some wide phonetic classes. Segmental parsing method was created. It allows to accelerate and greatly simplify the sentence parsing, which reduces the number of operations when searching for related words. The results of automated parsing are included in the integrated model as syntactic (word frequency, word-parasites) and psycho-semantic (stylistic language constructions, intonation emphasis in phrases) individual characteristics. Software modules were implemented by the authors to allow numerical investigations in current research area. The modules have been integrated in the software package "Automated workplace for forensics expert". It was later often used when working at economic agreements. Numerical evaluations have shown that the use of processing and interphoneme diphone models achieve a very significant gains, both in recognition speed and the amount of memory required. The results of the work are the improving of the speed and reliability of speaker identification and verification, the ability to perform an assessment of record identification, automatically choose the best noise cleanup method within minimal damaging of the signal, reducing the amount the training data, provided an opportunity of text-independent identification.