The problem on building context dependent model in continuous mandarin speech recognition in order to avoid coarticulatory effects is descussed in detail in this paper. On the basis of information theory

the distance metric of the traditional clustering algorithm is first studied

which is the divergence of the model distribution and the difference in entropy result from model merging or splitting. Then the clustering algorithm based on decision tree is presented

which makes full use of the phonological rules. The model obtained from it is easy to be generalized

and this method demonstrates especially better when many triphones emerge that are not covered in the training material. In addition

the clustering and training procedure is discussed. At last

the speaker independent large vocabulary continuous speech recognition experiment shows that

if the recognition material is different from the training material

the recognition model obtained from the decision-tree-based clustering algoyithm reduces the error rate by 7. 95 %. However

the recognition model obtained from the traditional merge algorithm reduces the error rate only by 2. 63 %.