Recently some fast methods (LAESA and TLAESA) have been proposed to find nearest neighbours in metric spaces. The average number of distances computed by these algorithms does not depend on the number of prototypes and they show linear space complexity. These results where obtained through vast experimentation using only artificial data. In this paper, we corroborate this behaviour when applied to handwritten character recognition tasks. Moreover, we compare LAESA and TLAESA with some classical algorithms also working in metric spaces. (C) 1998 Elsevier Science B.V. All rights reserved.