期刊名称：Journal of Theoretical and Applied Information Technology
印刷版ISSN：1992-8645
电子版ISSN：1817-3195
出版年度：2021
卷号：99
期号：5
语种：English
出版社：Journal of Theoretical and Applied
摘要：The article deals with the development of neural network means for analyzing a voice signal to recognize the speaker�s emotions. We have established the possibility of improving these means through the use of a convolutional neural network of the SqueezeNet type, which determines the necessity to assess the effectiveness of such use. We have also determined that it is possible to assess the efficiency of using the neural network model experimentally by means of indicators of recognition accuracy and duration of training. A software implementation of SqueezeNet has been developed, with a training sample formed, using the publicly available TESS database, consisting of samples of voice signals with 7 emotions for 2 users. Mel-frequency cepstral coefficients are used as the parameters characterizing a voice signal. Using computer experiments, we have found that after 80 periods of training on a fairly limited training sample, SqueezeNet enables using validation examples to achieve speaker recognition accuracy of about 0.95, which is proportionate to the results of the best modern systems of the similar purpose and confirms the possibility of effective use of this type of network for analyzing a voice signal. We have shown the necessity for further research related to the adjustment of neural network solutions to the recognition of the speaker�s emotions under a variety of noise interference. We have also determined the feasibility of developing a method for adjusting SqueezeNet architectural parameters to the conditions of the task to analyze a voice signal for simultaneous recognition of the speaker�s personality and emotions.
