Model pro prediktivní určování kvality hovoru jako kritéria pro hodnocení výkonnosti VOIP řešení

Abstract

The purpose of this thesis lies with the exploration of possible ways to detect, measure and estimate the quality of speech in IP networks based exclusively on protocol header information and real-time measurements. The knowledge uncovered by this exploration is used to form the basis of a generally usable model, which would implement the non-intrusive methodology for speech quality estimation. This model is meant to be used in the field of performance testing and benchmarking of telecommunication infrastructure based on both IP and SIP, where it is supposed to increase the precision and reliability of the commonly used methods. Moreover, the general nature of the model allows it to be used in monitoring systems as well, which is beneficial for the telecommunications service providers, for whom the model opens the way to proactively react on the speech quality degradation by the service routing modification or network parameters change. Throughout this thesis, the reader is familiarized with the network parameters affecting the speech quality and is presented with the neural network based model capable of estimating the speech quality based on the packet loss and its characteristics. This model is then integrated into the E-model system of estimating the effect of remaining relevant elements, so that it is possible to quantify the role of time characteristics as well. The model as the whole is a subject of simulations under various conditions and neural network settings and when the optimum is found the model is combined with the author-developed system for performance testing of SIP infrastructure. This way the possible usage of the presented model is verified. The reliability and accuracy of the model is verified for two selected codecs - G.711 A-law and SPEEX.