In this paper we investigate the minimum classification error (MCE) criterion for the training of distance measures for template based speech recognition. These MCE-based distance measures are illustrated with example experiments on the Wall Street Journal 5k benchmark for continuous speech recognition.