Abstract

This letter puts forward a supervised ML technique to determine the Quality of Experience (QoE) of VoIP calls. It takes its beginning from an investigation on VQmon, an enhanced E-model version that estimates the quality of IP-based voice calls adopting an objective approach. The current study demonstrates VQmon shortcomings via a comparison between the Mean Opinion Score (MOS) values this technique predicts and the actual average ratings collected from a subjective listening quality campaign. It proposes to deploy Ordinal Logistic Regression (OLR) for speech quality assessment, and results disclose that OLR outperforms popular ML algorithms, in accuracy and confusion matrices.

Highlights

  • Among ML classification algorithms, the Decision Tree classifier (DT) is a solution that produces interpretable models and it is widely employed for this distinctive feature: its goal is to create a model that learns from simple if/else rules inferred from data

  • The input features we gathered from the testbed in Fig.1 included the actual network metrics associated with each evaluated call, that is, the following numerical features: average and maximum jitter, number of received packets, packet loss rate, out-of-sequence packets and duplicated packets

  • To validate the goodness of such a choice, we considered a random split of the Quality of Experience (QoE) scores, employing 80% of them as the training set and the remaining 20% as the test set and first benchmarked Ordinal Logistic Regression (OLR) classification accuracy against that of the Random Classifier (RC), DT, Random Forest (RF) and Multinomial Logistic Regression (MLR)

Read more

Summary

Rationale and Contribution

Quality of Experience (QoE) of VoIP calls is a relevant topic in the realm of contemporary networks, given the widespread adoption of VoIP in wired scenarios. Taking the last remark as its foremost motivation, the aim of this letter is two-fold: (i) first, to quantify VQmon® limits in QoE assessment of VoIP calls that employ a wideband voice codec; (ii) to overcome such limits proposing the adoption of a supervised ML approach. With reference to the latter point, the current study demonstrates to what extent OLR performs better than other popular state-of-the art ML solutions. In the limiting case where ratings are collapsed on a coarse binary scale, OLR and alternative ML models are verified to guarantee a very high and comparable accuracy level

Related Work
BACKGROUND
Prediction Models
Experiment Setting and Design
Data Set Preprocessing
Exploratory Investigation and Performance Assessment
CONCLUSIONS
Full Text
Paper version not known

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call