Abstract: Up to this time, the major challenge to sign language recognition is how to develop approaches that scale well with increasing vocabulary size In this paper, an approach to large vocabulary, continuous Chinese sign language（CSL） recognition is presented, which uses subwords instead of whole signs as the basic units Since the number of subwords is limited, HMM based training and recognition of the CSL signal become more tractable and have the potential to recognize enlarged vocabularies Furthermore, the proposed method facilitates the CSL recognition when finger alphabet is blended with gestures About 2400 subwords are defined for CSL One HMM is built for each subword, and then the signs are encoded based on these subwords A decoder that uses tree structured network is presented Clustering of the Gaussians on the state, language model and N best is used to improve the performance of the system Experiments on a 5119 sign vocabulary are carried out, and the correct rate is over 90% for continuous sign recognition