We proposed a novel approach for the optimisation of over-complete decompositions from a WPT dictionary based on a multi-objective genetic algorithm (MOGA). The MOGA allows to maximise the classification accuracy while minimising the number of features. For the purpose of obtaining appropriate features for state of the art speech recognizers, a classifier based on hidden Markov models (HMM) is used to estimate the capability of candidate solutions, using on a set of English phonemes.