Quality of Experience Prediction for VoIP Calls Using Audio MFCCs and Multilayer Perceptron

Abstract

To provide a high-quality communication service to their users, VoIP service providers use some monitoring and warning systems that notify them of any malfunctions that may occur in the system. Because the VoIP service is delivered over the internet, issues with the internet infrastructure and related hardware have a direct impact on the quality of service (QoS) and experience provided. In such cases, service providers analyze the QoS reports to analyze the incidents. The QoS reports consist of various parameters such as packet loss, delay, jitter, and codec information extracted from the related VoIP call. However, in some cases, these parameters may be insufficient or corrupted. Therefore, real sound recordings are used to determine the source of the complaint. However, listening to audio recordings made by third parties is not preferred when the content is sensitive. Thus, a computer-based analysis is an important requirement in such cases. In this study, a machine learning-based model was developed that can classify a given packet loss into six classes, which is one of the most important factors affecting the quality of experience. The audio recordings were represented with Mel Frequency Cepstrum Coefficients (MFCCs). The model trained using 9000 5-second audio recordings from 15 different speakers can predict the packet loss rate and the mean opinion score (MOS) with an accuracy of 87%.