A new method to improve the signal-to-noise ratio of single evoked potentials (EP) measurements is presented, in which, contrary to previous methods, no a priori assumptions on the signal are necessary. This method is based on the wavelets decomposition of the individual signals. A statistical thresholding is applied on the coefficients of the decomposition: we estimate whether the mean value of the coefficients across trials and for each time point is significantly different from a random estimate. The performance of the method is evaluated with simulation and the method is applied to real data