Intelligent learning techniques applied to quality level in voice over IP communications

Abstract

This paper presents a method for determining the quality of a Voice over IP communication using machine learning techniques. The solution proposed uses historical values of network parameters and communication quality in order to train the different learning algorithms. After that, these algorithms are able to find the quality of the Voice over IP communication based on network parameters of a specific period of time. Intelligent and other machine learning algorithms take as input a baseline file that contains some values of network parameters and voice coding, associating an index quality for each scenario according to ITU-T Recommendation G.107. The tests were performed in an emulated network environment, totally isolated and controlled with real traffic of voice and realistic IP network parameters. The quality ratings obtained for the learning algorithms in all the scenarios were corroborated with the results of the algorithm of ITU-T Recommendation P.862. The results show the reliability of the four learning algorithms used on the tests: Decision Trees (J.48), Neural Networks (Multilayer Perceptron), Sequential Minimal Optimization (SMO) and Bayesian Networks (Naive). The highest value of reliability for determining the quality of the Voice over IP communications was 0.98 with the use of the Decision Trees Algorithm. These results demonstrate the validity of the method proposed.