In this paper, a new algorithm to construct a multi-template dictionary for character recognition is presented. It is a repetitive process of partitioning and estimating the distribution by observing both within-class distribution and between-class information. The effectiveness of the algorithm is shown by experimental results.