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

Abstract

The purpose of this paper is to determine a solution to estimate the quality of a signal of
using time domain signal information and machine learning algorithms in
an environment that simulates wireless networks using Voice over Internet Protocol (VoIP). The methodology
employed was divided into three stages, and degradations were initially applied in an environment that
simulated 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, algorithms
implemented 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 that
were saved, 272 of them were used for training on 12 different learning algorithms.
implemented in the Weka tool. In the last step, software-trained algorithms
implemented in Java called PredictorFX in order to predict the value of MOS through
an audio image in the time domain. The results were satisfactory, the best
trained regression algorithms called r1 were RandomTree, RandomForest and IBk with
correlation coefficients ranging from 0.9798 to 0.9982 in the validation phase. In relation to r2 the
best were RandomTree, RandomForest, IBk and AditiveRegression with correlation coefficient
ranging from 0.9375 to 0.9923 in the validation phase. And finally, for the training algorithms for the
named c1 classification the best trained algorithms were IBk, RandomTree, RandomForest
and J48 with a range of 48.53% to 98.53% of correctly classified instances.