Dimension reduction is often included in pattern recognizers based on hidden Markov models to lower the size of the models to estimate. Commonly used methods are heuristic in nature and do not take care of information retention after projection. In this paper, we present a new method based on the approach of suficient dimension reductions. It explicitly accounts for all the discriminative information available in the original features, while using a minimum number of linear combinations of them. We review the underlying theory and present an algorithm for the practical implementation of the proposed method. On the experimental side, we use simulations to illustrate its advantages over widely-used existing alternatives. In particular, we show that it performs as good as existing techniques when data is optimal according to the assumptions of those techniques, but signi cantly better for heteroscedastic data with no special structure on the covariance matrix.Sociedad Argentina de Informática e Investigación Operativa