Applications rely on rich multimedia contents and experience of end users is sensitive to network conditions. Consequently, network operators must design their infrastructure to ensure high Quality of Experience (QoE) for their customers. However, applications are usually over-the-top services on which network operators have no control and users have no mean to tune the network when they undergo poor QoE. In this paper, we propose a method that allows network operators to determine how their network performance will influence QoE and end users to predict the QoE even before launching their applications. We predict the subjective QoE users will undergo based on the knowledge of objective network performance parameters obtained with active measurements (e.g., delay, loss) and machine learning. With the particular case of Skype calls and using a decision tree, we show that our approach achieves 83% of accuracy when estimating QoE from the delay, bandwidth, and loss. Our method can be seen as a new way of performing measurements at the Internet access, where instead of expressing the expected performance in terms of network-level measurements, the performance of the access is expressed in clear terms related to the expected quality for the main applications of interest to the end user. The strength of the approach is in its capacity of expressing directly the QoE as a function of network-level measurements, which is an enabler for QoE prediction, and in reusing the same network-level measurements as input to different models for the QoE of end user applications.