In template based speech recognition, there is a need for a high-performant distance measure between speech frames. Some well known metrics include the Euclidean and the Mahalanobis distance. The recent tendency is to perform a local scaling of the distance metric, defining a set of classes and computing a set of weights for each of these classes. Discriminative training approaches have already proven their usefulness in various domains including speech recognition. They have the well known characteristic of training the weights for all of the classes simultaneously, and not independently of each other. In this paper, a first attempt is made to incorporate a discriminative distance measure into template based speech recognition. We use a distance measure trained by a very intuitive discriminative criterion and show that it works very well, even beating the performance results of comparable HMM-based speech recognizers.