This paper explores the use of wavelet packet analysis to extract features from spontaneous electroencephalogram (EEG) during three different mental tasks. Artifact-free EEG segments are transformed to multi-scale representations by dyadic wavelet packet decomposition channel by channel. Their feature vectors formed by energy values of different sub-spaces EEG components are used as inputs of a radial basis function network to test the classification accuracies of three task pairs. The results indicate that the classification accuracies of the wavelet packet analysis method are significantly better than those of autoregressive model method. Wavelet packet analysis would be a promising method to extract features from EEG signals.