Determining a Non-Intrusive Voice Quality Model Using Machine Learning and Signal Analysis in Time

• Autor(a) principal: Zegarra Rodriguez, Demostenes
• Data de Publicação: 2019
• Outros Autores: Brandão Junior, Luiz Carlos
• Tipo de documento: Artigo
• Idioma: eng
• Título da fonte: INFOCOMP: Jornal de Ciência da Computação
• Texto Completo: https://infocomp.dcc.ufla.br/index.php/infocomp/article/view/630

The purpose of this paper is to determine a solution to estimate the quality of a signal ofusing time domain signal information and machine learning algorithms inan environment that simulates wireless networks using Voice over Internet Protocol (VoIP). The methodologyemployed was divided into three stages, and degradations were initially applied in an environment thatsimulated wireless networks making changes in two parameters being the signal-to-noise ratio (SNR)and the type of modulation scheme. To perform the degradations on six distinct signals, algorithmsimplemented in MATLAB were used to simulate the effect of fading in wireless environments.In the second step, time domain graphs were plotted that correspond to the degradations and thatwere saved, 272 of them were used for training on 12 different learning algorithms.implemented in the Weka tool. In the last step, software-trained algorithmsimplemented in Java called PredictorFX in order to predict the value of MOS throughan audio image in the time domain. The results were satisfactory, the besttrained regression algorithms called r1 were RandomTree, RandomForest and IBk withcorrelation coefficients ranging from 0.9798 to 0.9982 in the validation phase. In relation to r2 thebest were RandomTree, RandomForest, IBk and AditiveRegression with correlation coefficientranging from 0.9375 to 0.9923 in the validation phase. And finally, for the training algorithms for thenamed c1 classification the best trained algorithms were IBk, RandomTree, RandomForestand J48 with a range of 48.53% to 98.53% of correctly classified instances.