Novel speech features purposely designed for speaker verification are proposed. Based on the selec- tion of the most representative basis over all possible bases of the wavelet packet transformed signal, this representation leads to optimal speech features fine- tuned for differentiation of human voices. Our method can be easily extended to other classification problems and can be used with other libraries of orthonormal bases. The practical significance of our approach has been evaluated in comparative experiments performed on the Polycost speaker recognition database. The pro- posed speech features demonstrated superior perform- ance when compared to other wavelet packet-based speech features, and to the widely-used Mel-frequency cepstral coefficients (MFCC).