This paper proposes a novel modular initialization scheme of Multilayer Perceptrons (MLPs) trained for phoneme classification. Small MLPs are trained to discriminate between a phoneme and all the others. In the next step they are merged using our novel initialization scheme in broad classes and trained further. In the last step we merge the broad phonetic MLPs using the same scheme to generate the final phonetic MLP. Experiments done on a Dutch language isolated word database showed that the scheme gives faster and better estimates of Bayesian a posteriori probabilities compared to random initialization. Moreover, given its modularity, the method offers the possibility to deal with high dimensional problems.