This paper describes methods developed to improve the performance of a preselection algorithm in a large vocabulary isolated word recognizer. First we investigate ways for optimizing the phonetic representations of the words in the lexicon (the base forms of which are obtained from a grapheme-to-phoneme converter) in such a way that the inherent limitations of the front-end are taken into account and where possible are remedied. Then an attempt is made to improve the quality of the symbol strings produced by the acoustic front-end. keywords: Isolated-Word-Recognition preselection lexicon