COMPARATIVE ANALYSIS OF THE EFFECTIVENESS OF NEURAL NETWORKS AT DIFFERENT VALUES OF THE SNR RATIO

Abstract

This work is devoted to a comparative analysis of the effectiveness of neural networks, CNN and RNN, at different SNR ratios.  The research conducted within the framework of this work showed that CNN convolutional neural networks demonstrate higher efficiency in speech signal recognition tasks, regardless of different levels of SNR ratio and language. Thus, the CNN neural network showed stable superiority over RNN under all conditions, especially at low SNR ratios. It was revealed that with an increase in the SNR ratio, the difference in accuracy between the CNN and RNN neural networks decreases, but the CNN continues to lead, which indicates its higher adaptability and ability to learn under conditions of different noise and interference levels. It is especially important to note that the advantage of CNN becomes noticeable at low SNR values, where the accuracy of the RNN decreases more significantly. As a result, with an SNR ratio of 3 dB, the recognition accuracy using CNN was 80% for the Kazakh language, whereas RNN showed a result in the region of 75%. With an increase in the SNR ratio to 21 dB, the difference in accuracy between CNN and RNN decreased, but CNN continued to lead, reaching 88% accuracy compared to 86% for RNN. In addition, the results showed that the effectiveness of the CNN and RNN depended on the language in which they were trained. Neural networks trained in Kazakh showed the best results in recognizing Kazakh speech but also successfully coped with recognizing the Russian language. This highlights the importance of considering language features when developing and training neural networks to improve their performance in multilingual environments.