Introduction

This Chapter proposes a new method to evaluate the quality of real-time multimedia streams, eventually transmitted over packet networks (IP or ATM networks). We present the methodology in general, which is applicable for whatever media type (speech, audio, video or multimedia). In addition, real-time application configuration such as session types (one way, interactive two ways, or multiparty conferences) can be taken into consideration. The validation of this method is left to the next subsequent Chapters. It is known that subjective quality measures are the most reliable methods because, by definition, they are carried out by human subjects; that is, their results correspond to human (end-users) perception. However, they are very difficult to carry out, expensive and time-consuming tasks. Furthermore, they are not suitable to nowadays real-time applications running over packet networks (e.g. VoIP, videoconferencing, etc.). On the other hand, objective quality methods can be, in certain cases, used to evaluate the quality in real-time, as they are computer programs. But their major disadvantage is that their results do not always correlate well with human perception (subjective quality measures). It is known that media-type quality is affected by many parameters (we name them as the quality-affecting parameters), such as the network loss rate, the delay variation, the encoding type, the bit rate, etc. In addition, the quality is not linearly proportional to the variation of any of these parameters. The determination of the quality is, therefore, a complex problem, and it has not been possible to solve by developing mathematical models that include the effects of all these parameters simultaneously. Our problem has two aspects: first, a classification one, the mapping between the parameters' values and the quality; second, a prediction one, the evaluation of the quality as a function of the quality-affecting parameters in an operational environment. We believe that neural networks (NN) are an appropriate tool to solve this two-fold problem [117,19]. We illustrate our approach by building a system that takes advantage of the benefits offered by NN to capture the nonlinear mapping between several non-subjective measures (i.e. the quality-affecting parameters) of media-type sequences transmitted over packet switched networks and the quality scale carried out by a group of humans subjects. Hence, a suitable NN is used to learn the mapping between the quality measures and the parameters' values. The NN then can be used to automatically predict the quality for any given combination of the quality-affecting parameters. The rest of this Chapter is organized as follows. In Section 4.2, we provide a general description of our method. After that, we discuss in Section 4.3 the subjective quality tests to be used with our method for each type of media. We focus in Section 4.4 on the mean opinion score calculation. As our new method is based mainly on neural networks (NN), we present in Section 4.5 their use in our method, as well as a comparison between the two models of NN used (ANN and RNN). The goal of Section 4.6 is to overview the different quality-affecting parameters for multimedia types that can be used with our method. Some possible uses and application of our method are outlined in Section 4.7. Section 4.8 is on the run-time mode of our method when integrated to measure in real time multimedia quality. Finally we provide the conclusions of this Chapter in Section 4.9