In this paper, we present a new approach to unconstrained, off-line handwriting recognition (HWR). It is based on the global plausibility estimation of a word knowing the local probability of each individual character. Having a set of character models issued from a training step and an input word pattern, we maximize for each word of the lexicon the sum of plausibilities of its component characters. This maximisation is made in terms of observation quality and extent of a symbol within the pattern. The proposrd method operates in a top-down manner by giving srgmentation hypotheses which induce local symbol rxtcnts for a given word of the lexicon against which the pattern is matchrd. The word which obtains the highest average plansihility per character is labeled as the recogni7rtl onr. This method was applied with success on continuous sprrch rrrognition by [5]