In this paper the probabilistic aproach to handwritten words recognition is described. The decision is performed using results of character classification based on a character image analysis and probabilistic lexicon treated as a special kind of soft classifier. The novel approach to combining these both classifiers is proposed, where fusion procedure interleaves soft outcomes of both classifiers so as to obtain the best recognition quality. The proposed algorithms were experimentally investigated and results of recognition of polish handwritten surnames and names are given.