Abstract: Context-dependent acoustic model based on decision tree has been deeply investigated and applied in western language speech recognition.But in Mandarin speech recognition,diphone model was more popular and little attention was paid to triphone in the past.In this paper triphone model based on decision tree was proposed.and some key problems are discussed which must be solved when build triphones,such as how to choose the basic model unit set,how to design the question set,how to choose evaluation function,the choice of stop criterion,and how to build the decision tree and triphone model.The experiments showed that error rate was reduced by 24.7% in continuous speech recognition system based on triphone compared with one based on diphone model.