Quality of experience prediction for VoIP calls using MFCCS and machine learning algorithms

The advent of technological advancements has facilitated the emergence of internet infrastructures, thereby enabling convenient access to communication services that boast both high quality and affordability. On the other hand, hardware and software problems directly affect the quality of service offered and reduce the experience quality of users. In call centers where a high number of calls are made at the same time, these problems should be detected before a complaint is received. For this, the source of the problem should be quickly identified by monitoring and reporting tools. In this study, Random Forest (RF), Multilayer Perceptron (MLP), Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), and Consistent Rank Logits (CORAL) were used to estimate the communication quality experienced by call center users on Unified Communications as a Service (UCaaS) infrastructure. Predictions were made in 6 classes (No Loss, 5% Loss, 10% Loss, 15% Loss, 20% Loss and 25% Loss). The obtained results showed that the five different models, namely MLP, CORAL, CNN, RF, and LSTM, achieved accuracies of 86.4%, 85.3%, 84.8%, 82.9%, and 78.6%, respectively. On the other hand, the highest accuracy rate of 98.4% achieved in the No Loss class with MLP, while the lowest accuracy rate of 61.2% achieved in 20% Loss class with the LSTM model. Except for MLP, all models showed a declining trend in accuracy rates from the No Loss to 20% Loss classes.