We describe a statistical frequency domain approach to localizing equivalent dipole generators of human brain evoked potentials. The frequency domain representation allows considerable data reduction, constrains the magnitude function of the dipoles to be smooth, and accounts for the statistical properties of the background EEG. A previous paper described a restrictive model in which the dipole orientations were assumed to be fixed over time, and only one dipole was allowed. In this paper, we consider the more general model in which the orientation can vary over time, and which includes multiple dipole generators. The varying orientation model has the practical advantage of being more nearly linear and more flexible than the fixed orientation model, which facilitates convergence of the iterative fitting algorithm. We suggest a measure of goodness-of-fit that compares the likelihood of the dipole model with the likelihoods of saturated and null models. We report the results of fitting the model to recorded auditory and visual evoked potentials. A single dipole with fixed orientation seems to be an adequate model of the auditory midlatency response, while two dipoles with varying orientation are needed to fit the later P200 component. Analysis of the visual P100 response to unilateral stimulation localized a generator in the contralateral occipital cortex, as expected from anatomical considerations. A two-dipole model fit the visual P100 response of bilateral stimulations, and the locations of the two dipoles were similar to the locations obtained by single-dipole fits to the responses to left and right unilateral stimuli.