In this paper, a real-time system designed for recognizing continuous Chinese Sign Language (CSL) sentences with a 4800 sign vocabulary is presented. The raw data are collected from two CyberGlove and a 3-D tracker. The worked data are presented as input to Hidden Markov Models (HMMs) for recognition. To improve recognition performance, some useful new ideas are proposed in design and implementation, including states tying, still frame detecting and fast search algorithm. Experiments were carried out, and for real-time continuous sign recognition, the correct rate is over 90%.