Akat'ev

This article deals with the problem of variability of word pronunciation in analysis and speech recognition tasks. An adaptive acoustic model defined as a multitude of minimal sound units (MSU) united into a cluster-phoneme under the principle of minimum informational mismatch in Kullback-Leibler metric, is proposed. An adaptive algorithm of filling the MSU cluster from a continuous stream of speech was developed on the basis of the whitening filter method. An example of its practical implementation is also provided in the article. As a result of this experiment, from the total list of phonemes of the national language the authors selected the phonemes which, in their implementation, are the most sensitive to conditions of their pronunciation by the speaker. Adjusting an information system to such a phoneme, the authors guarantee maximum sensitivity of perception in relation to the speaker’s emotional and physical state.