Same but different? - Using speech signal features for comparing conversational VoIP quality studies

Measuring speech quality in Voice over Internet Protocol (VoIP) networks is an increasingly important application for legal, commercial and technical reasons. Any proposed solution for measuring the quality should be applicable in monitoring live-traffic non-intrusively. The E-Model proposed by the International Telecommunication Union-Telecommunication Standardisation Sector (ITU-T) achieves this, but it requires subjective tests to calibrate its parameters. In this paper a solution is proposed to extend the E-Model to any new network conditions and for newly emerging speech codecs without the need for the time-consuming, expensive, hard to conduct subjective tests. The proposed solution is based on an artificial neural network model and is compared against the E-Model to check its prediction accuracy.