Abstract: In this paper we investigate three different approaches for the global modeling and recognition of the words used to write the legal amount on French bank checks (27 lexicon entries), mainly written in mixed cursive and discret style. The first Model is a global one since it does not require any explicit letter level. The second Model is built to use the explicit concatenation of letter Models and is called ``a letter reconstruction based approach''. The third Model is able to give each grapheme its corresponding interpretation within a word (either part of a letter, letter or group of letters) and has been called a grapheme reconstruction based approach. To analyse the three approaches independently from a specific description, each of them uses the same segmentation process and feature set. The three approaches have been tested on real images of Bank Checks scanned for the French Postal Technical Research Service (SRTP).