Despite a large literature in Quality of Service (QoS) evaluation, end-user QoS provisioning remains an open research field. There is no general agreement about what to measure and how to do it in order to ensure real quality levels. Even more, new heterogeneous multimedia applications have redefined the problem, turning many previous implementations no longer appropriate for current scenario. This paper addresses the problem of QoS assessment of a multimedia service over IP as perceived by humans, applying statistical learning techniques. We describe two end-to-end performance evaluation methodologies, the former based on Perceived QoS (PQoS) and the latter based on functional nonparametric regression. By merging them we build an improved system for end-to-end PQoS evaluation which allows analysing and better understanding the tradeoffs between different proposed techniques in the field.