Neurological signals are generally very weak in amplitude and strongly noisy. As a result, one of the major challenges in neuroscience is to be able to eliminate noise and thus exploit the maximum amount of information contained in neurological signals (EEG, ERP, Evoked Potentials, ...). In our project, we aim to highlight the ERP's N400 wave which the behavior, the amplitude and the latency may reflect the effects of vowelling and semantic priming in Arabic language. For that reason, we consider a nonlinear filtering method based on discrete 10th order Daubechies discrete wavelet transform combined to principal component analysis, to improve the quality of the recorded ERP signals. Thus, among all tested wavelets, the Daubechies one allows a significant improvement of the signal to noise ratio while using only 10 ERP trials. In addition, we compare and illustrate the effectiveness of this method to that obtained using the averaging technique implemented on EEGLab toolbox. In a second step, the Mexican Hat function have been used to achieve continuous wavelet analysis of the filtered signals. This method permits us to get an alternative representation of the ERPs and to detect avec more accuracy the N400 wave.