In this paper, we focus on spectral parameter filtering for reducing the mismatch between training and testing and report experimental results on a continuously spelled name recognition task over the telephone. We studied various time derivative feature combinations, the influence of RASTA processing, short-term and long-term cepstrai mean normalization, and the influence of the amount of training data on recognition performance. Based on the results of these experiments, we derived a new front-end for our task, leading to an error rate improvement of almost 30% in name retrieval as compared to previous published results. We also discuss the interaction between the different techniques studied when used in combination.