Traditionally, the problem of cursive script recognition has been handled in two fundamental ways: one based on the global approach and the other on the segmentation approach. In this paper, we present an inexact segmentation approach to segment the word into letters or even strokes. Our algorithm searches for high curvature points along the lower contour of the image profile. These points are then treated as segment points and marked as potential letter boundaries. Using the histogram profile, an initial estimation of the word length is made. Yet, this estimated word length may be adjusted later. In the process, dynamic programming is used as a general optimization technique to produce a list of possible candidates of the same word length. If the score of these candidates cannot meet some recognition criteria, the word length is re-estimated and another set of candidates are generated. Eventually, an entropy- based measure is provided for comparison among candidates of different word length. In our system, a contextual postprocessor can easily be added to further improve the recognition rate. But in this instance, we have an additional advantage in that a large number of candidates of variable word length are available for selection and even some words with unrecognized letters may also be taken into consideration. In this paper, some experiments are also carried out to evaluate the performance of our novel algorithm.