简介：Toefficientlydecreasethesizeofparametersandimprovetherobustnessofparameterstraining,afuzzyclusteringbasedphonetictied-mixturemodel,FPTM,ispresented.TheGaussiancodebookofFPTMissynthesizedfromGaussiancomponentsbelongingtothesamerootnodeinphoneticdecisiontree.FuzzyclusteringmethodisfurtherusedforFPTMcovariancesharing.ExperimentalresultsshowthatcomparedwiththeconventionalPTMwithapproximatelythesameparameterssize,FPTMdecreasethesizeofGaussianweightsby77.59%andincreaseswordaccuracyby7.92%,whichprovesGaussianfuzzyclusteringisefficient.ComparedwithFPTM,covariance-sharedFPTMdecreasesworderrorrateby1.14%,whichprovesthecombinedfuzzyclusteringforbothGaussianandcovarianceissuperiortoGaussianfuzzyclusteringalone.