This paper proposes a new and accurate way of predicting the end-to-end performance of a multimedia stream. It consists of coupling an approach we have previously developed, which is able to capture very precisely the way humans assess an arriving stream, with classical performance evaluation models. The former approach can automatically quantify the quality of the connection as humans would do (using statistical learning tools), considering this quantitative measure of quality as a function of measurable parameters of the network and the source; the latter allows mapping this "ultimate" quality value to input parameters such as offered traffic or loads, through, for instance, queuing models. In the paper, we also propose a simplification of the global approach dealing with simple closed formulas. We illustrate our approach with a simple case study of unidirectional VoIP performance.