Parameter-tying (or sharing) is widely used in hidden Markov models (HMM) for speech recognition because of its ability to enhance recognition accuracy and robustness. This paper tries to make best use of phonetic structure of the Chinese language to optimize parameter-tying in building acoustic models for an HMM-based continuous Chinese speech recognition system. Phonetic context based parameter-tying schemes are studied and compared with conventional parametertying schemes through experiments on a very large vocabulary, speaker-independent, continuous Chinese speech recognition system. The test of four male and four female subjects shows that the proposed parametertying scheme gives substantial improvement over the conventional ways of parametertying, without significantly increasing the number of system parameters.