Possibilities of Using Neural Networks to Assess Speech Intelligibility

The paper is devoted to the consideration of modern speech recognition technologies for speech intelligibility evaluation. Speech intelligibility is an important criterion for evaluating the quality of communication lines in various fields (telephony, radio, information security). Currently, to assess speech intelligibility in the tasks of information security use an instrumental-calculative method based on the results of experimental research conducted by N.B. Pokrovsky. The methodology of N.B. There is also a more accurate way to assess speech intelligibility - the method of articulation. The criterion of intelligibility here is the intelligibility of speech elements (sounds, syllables, words or phrases) - a value defined as the ratio of the number of correctly received speech elements to the total number of transmitted elements. However, this method has certain disadvantages, in particular, the cumbersomeness of articulation measurements, which require an articulation team and a large amount of time. As a result, the question arises whether it is possible to eliminate these disadvantages by replacing the receiving operator with an artificial equivalent. It is necessary to experimentally evaluate the performance of modern speech recognition technologies in recognizing noisy speech with different signal-to-noise ratios.