Algorithm development for recognizing human emotions using a convolutional neural network based on audio data

Abstract

Objectives. This article provides a description and experience of creating the algorithm for recognizing the emotional state of the subject.Methods. Image processing methods are used.Results. The proposed algorithm makes it possible to recognize the emotional states of the subject on the basis of an audio data set. It was possible to improve the accuracy of the algorithm by changing the data set supplied to the input of the neural network.The stages of training convolutional neural network on a pre-prepared set of audio data are described, and the structure of the algorithm is described. To validate the neural network different set of audio data, not participating in the training, was selected. As a result of the study, graphs were constructed demonstrating the accuracy of the proposed method.After receiving the initial data of the study, the analysis of the possibilities for improving the algorithm in terms of ergonomics and accuracy of operation was also carried out. The strategy was developed to achieve a better result and obtain a more accurate algorithm. Based on the conclusions presented in the article, the rationale for choosing the representation of the data set and the software package necessary for the implementation of the software part of the algorithm is given.Conclusion. The proposed algorithm has a high accuracy of operation and does not require large computational costs.