Determinação de um modelo não intrusivo de qualidade de voz fundamentado na análise do sinal no domínio do tempo usando aprendizagem de máquina

Abstract

Voice over Internet Protocol (VoIP) is one of the communication services that emerged in the early 1990s. In recent years, the capacity of IP networks has increased, and technology has gained more space by performing investment in quality of service. In this work, a solution is proposed to estimate the quality of a voice signal using signal information in the time domain and with the support of machine learning algorithms. The methodology was divided in three stages. In the first one, degradations were applied in environments that simulated wireless networks, making changes in two parameters that were, the signal-to-noise ratio (SNR) and the type of modulation scheme. In the tests, six different original sound signals were used. To perform these degradations, algorithms implemented in MATLAB were used to simulate the effect of fading in wireless environments. In the second step, graphs of the degraded audio signals were written, in the time domain that were saved, 272 images were used to train in 12 different machine learning algorithms implemented in the Weka tool. In the last step, the trained algorithms were placed in a Java-based software called PredictorFX in order to predict the value of MOS using an audio image in the time domain. The results were satisfactory, the best Regression Algorithms (ATR) were RandomTree, RandomForest and IBk with their correlation coefficients varying from 0.9886 to 0.9989 in the validation phase for the data that resulted in the MOS, called trained regression algorithm (ATR1). In relation to ATR2, which contains the information extracted from the images, the best algorithms were RandomTree, RandomForest, M5P and MLP, with correlation coefficient varying from 0.8638 to 0.9896, in the validation phase. Finally, for the Classification Training Algorithms (ATC) called ATC1, the best algorithms were OneR, J48, MLP and RandomForest with 58.82 % to 96.32 % of the correctly sorted instances. These results demonstrate that it is possible to conduct non-intrusive voice quality tests using models based on the ITU-T Recommendation P.862.