Multi-agent Algorithm Imitating Formation of Phonemic Awareness

Abstract

This paper proposes the cognitive speech perception model necessary as a theoretical basis for the development of universal automatic speech recognition systems that are highly effective in conditions of high noise and cocktail party situations. A formal description of the general structure of the act of speech perception and the main elements of the structural dynamics of the speech recognition process has been developed. The necessity of using the articulation event as a minimal basic pattern of sound image recognition has been proved. Using articulation event gives an opportunity to analyze such aspects of speech message as extra-linguistic components and intonation means of expression. Multi-agent systems are chosen as the formal means of implementation. An algorithm for supervised machine learning with an imitation of the mechanism of the formation of a human’s phonemic awareness is developed. It gives the possibility to create speech systems that are resistant to the diversity of accents and individual characteristics of the user.