This paper proposes a new method for connected word recognition. To reduce pattern matching errors which are caused by coarticulation, talking rate variation, and silence between words in continuous speech, overlap and splitting of reference patterns and normalization of accumulated distances are introduced into our basic recognition algorithm. Also new algorithms to reduce the amount of computation and to train the recognition system are proposed. Performance evaluation tests show the efficiency of these algorithms and the entire method. Dominant factors of speech data influencing recognition performance are also investigated.