We propose a new approach for optimally pre-enhancing speech signals for given noise conditions. Like others, we optimise the predicted intelligibility of the signal, however, we employ a statistical `microscopic' intelligibility model that encodes information about which spectro-temporal speech regions are most informative. Uniquely, our optimisation strategy aims to maximise the discrimination between the correct interpretation and competing incorrect interpretations of the utterance. We present results from studies that use speech-shaped stationary noise maskers and show the new strategy leads to solutions that are more varied than the simple high frequency emphasis employed in many pre-enhancement systems.