In this paper we discuss a non-invasive method to localize neural electrical sources using EEG data. In this method, the human head is modelled by a set of four concentric spheres with different conductivities which represent the scalp, the skull, the CSF and the brain or by three triangulated surfaces which approximate the exact head shape (in this model we do not consider the CSF layer) using NMR images. In this case the computer effort is very high, since the calculations imply thousands of equations. Therefore, the number of research groups working with this improved model, in the world, is very small. In both models, we assume that the neural source is a current dipole. This makes the model suitable for cases where the active brain areas are limited and localized. We discuss some error factors associated with the method, as the geometry of the head, the conductivity of the different layers and the number of electrodes used in the EEG measurements. Comparing the more realistic head model, with the spherical one we often have differences of 1-2 cm. However, we can reach even more pronounced differences in the frontal areas. Concerning the skull conductivity, we realized that it could introduce errors of 1-2 cm. We observed that at least 50 electrodes should be used only since 21 electrodes could imply errors of about 0.5 cm. The method was applied, both in the spherical version and the realistic one, to clinical cases of focal epileptic patients. The results are discussed in terms of the other clinical information available and they are coherent with the remaining clinical data.