Extraction of stimulus evoked potentials (EP) out of the data recorded at the scalp is a difficult problem because of poor signal-to-noise ratio. Ensemble averaging techniques do not always give entirely satisfactory results because of signal variability, in terms of shape and latency variations. To cope with this signal variability many signal processing methods have been proposed. These include Wiener filtering and its extensions and recently, Neural Network signal processing techniques. Here we use an approach in which two different, parametrically- described models are considered for the spontaneous and the evoked parts of the measured activity. The model parameters are identified using a special formulation which converts the identification problem into a (nonlinear) filtering problem. Extended Kalman Filtering (EKF) technique is thus used for the identification of model parameters. Once the model parameters are obtained, Kalman filtering is used once more to obtain an estimate of the evoked part of the signal. Some modifications to the EKF algorithm have been incorporated in order to overcome divergence problems associated with the Extended Kalman Filter.