In the current conditions of computational linguistics rapid development and competing demands of the natural language automatic processing the issue of successful speech recognition is becoming increasingly important opening up new opportunities for the effective implementation of the natural language interface.

The primary aim of effective speech recognition is accurate and immediate reproduction of the spoken speech signal into the recognized message in the context of the algorithm of classification (Dobrushkin & Danylov, 2009, p. 51). The recognized message can be presented both in a form of the text of the given message and in a convenient form immediately converted for its further processing in order to initiate the appropriate response of the software system (Yalova, Yashyna & Vasylieva, 2018,p. 102).

In general, effective speech recognition system should involve such interdependent parameters of complete input signal processing as denoising, segmentation, and determination of the vocalized areas, parameterization, recognition, and correction by the inverse dictionary (Dobrushkin & Danylov, 2009, p. 52). Taking into consideration the approaches applied nowadays, the solution of the abovementioned problem comprises a consistent comparison with the standards set by the vocabulary of the language recognition system.

Currently, the process of language recognition includes the following stages: recording and digitization of the speech fragment, calculation of temporal and spectral characteristics of the processed samples, classification of time frames by phonetic categories, selection of appropriate categories of lexemes, measuring the degree of reliability (Mykoliuk, 2018, p. 162). In the process of language recognition the most challenging issue is to carry out the procedure of comparing the input and reference element, the problem is further complicated by the fact that the input signals are characterized by certain temporal length (Yalova et al., 2018, p. 103).

Automated language recognition systems can be classified according to certain characteristics: language specifics, number of speakers, type of the structural unit (phrases, words, phonemes, diphons, allophones), principles of structural units separation, volume and completeness of the vocabulary to be recognized (Dobrushkin & Danylov, 2009, p. 53; Yalova et al., 2018, p. 104). Modern systems of speech recognition are developed on the basis of principles of forms recognition establishment. Among basic methods and algorithms of the automatic recognition of speech the following ones are presented: method of recognition of speech signals on the basis of signs (phonetic, non-acoustic, parametric), method of spectral entropy, correlation analysis, linear models, Hidden Markov Models, dynamic programming - Dynamic Time Warp (DTW), artificial neural networks, methods based on wavelet transformation and the Fourier transform, methods of the computational learning apparatus - artificial immune systems.

Neural networks, Hidden Markov Models and Dynamic Time Warp are considered to be among the most effective and advanced methods of speech recognition at the present stage.

Neural network approach. A neuron is a cell of the network that can appear in an excited or inhibited state. It has connections with such neurons in the network as synapses (the unidirectional input connections) and axons (the output connections of the neuron) through which signals (of excitation or inhibition) come to the synapses of the subsequent neurons. Using neural networks to recognize speech signals also involves building up an appropriately designed network and selection of the relevant weighting coefficients of synapses to minimize the errors (Yalova et al., 2018, p. 105).

Hidden Markov Models. The algorithm of Hidden Markov Models allows considering language as some specific code whereas the language stream is viewed as a sequence of some code packages. The element of such code (phonemes, syllables or even whole words) is not significant; the only thing that matters is that the probability of appearance of any element of the code depends on a number of previous elements. Thus, the language is generated by the Markov source and the language code is random (Mykoliuk, 2018, p. 162; Miasishchev, Loza & Bohachko, 2019, p. 133).

Dynamic Time Warp. The pronunciation of the same word usually has a different duration. Even if a word has been pronounced with the same duration, the pronunciation of certain parts of the word may vary. Therefore, in order to obtain an assessment of the difference between two speech signals in the form of a vector a time alignment realized by means of DTC must be performed (Yalova et al., 2018, p. 106).

Nevertheless, despite significant achievements in speech recognition initiated by modern computational technologies, the most complicated aspects of developing effective speech recognition systems still remain the following: “algorithms for reproducing the sequence of pronounced words, determination of an effective invariant acoustic model and formation of language models for certain languages” (Dobrushkin & Danylov, 2009, p. 50) which, in its turn, requires intensive researches in various fields of modern science: acoustic engineering, applied linguistics, neurolinguistics, corpus linguistics, computational linguistics.

References

1. Dobrushkin, H., & Danylov, V. (2009). Osnovni pidkhody do rozpiznavannia movlennievoi informatsii [Basic approaches to speech information recognition]. Informatsijni tehnolohiji ta kompiutery [Information technology and computers], 4, 50-64. [in Ukrainian].
2. Miasishchev, O., Loza, V., & Bohachkov, S. (2019). Doslidzhennia metodiv ta alhorytmiv rozpiznavannia movlennia [Research of speech recognition methods and algorithms]. Molodizhna viiskova nauka u Kyivskomu natsionalnomu universyteti imeni Tarasa Shevchenka: materialy Vseukrainskii naukovo-praktychnii internet- konferentsii molodykh vchenykh, adiunktiv, slukhachiv, kursantiv i studentiv (23 kvitnia 2021 r., m. Kyiv). [Youth military science at the Taras Shevchenko Kyiv National University: materials of the All-Ukrainian scientific-practical conference of young scientists, adjuncts, listeners, cadets and students (April 23, 2021, Kyiv)], 133. [in Ukrainian].
3. Mykoliuk, I. (2018) Analiz metodiv rozpiznavannia movlennia [Analysis of speech recognition methods]. XLVIINaukova ta tehnichna konferentsiia VNTU(14-23 bereznia 2018 r., m. Vinnytsia). [XLVII Scientific and Technical Conference of VNTU divisions (March 14-23, 2018, Vinnytsia)]. Retrieved June, 15, 2021, from https://conferences.vntu.edu.ua/index.php/all-fitki/an-fitki-2018/paper/view/5144 [in Ukrainian].
4. Petrov, A. (2012). Modeli ta metody rozpiznavannia movy [Models and methods of speech recognition]. Suchasnyi iformatsiinyi zahyst [Modern information protection], 1, 25-33. [in Ukrainian].
5. Yalova, K., Yashyna, K., & Vasylieva, A. (2018). Doslidzhennia system ta metodiv rozpiznavannia movnoho syhnalu [Research of systems and methods of speech signal recognition]. Tehnichi nauky [Technical sciences], 1, 101-106. [in Ukrainian].