Abstract: State duration in HMM is modeled by using time dependent state transition probability.Firstly,the traditional HMM is compared with the modified HMM that has incorporated duration information.Almost the same recognition performance is obtained by both HMMs for trained speaker in multi speaker mode.However,it is found that the modified HMM is more robust than the traditional HMM for untrained speaker in multi speaker mode.So it can be considered that the modified HMM that has incorporated duration information contains more phonetic transition information about the syllable to be recognized.Secondly,a new feature is proposed,which is called KLCEP,and an attempt has been made about how to combine various features such as LPCCEP,ARCEP,and KLCEP.ARCEP has been shown to be more robust for untrained speaker too and KLCEP is helpful to improve the performance of trained speaker.Finally,a high performance is obtained by using LPCCEP+ARCEP+KLCEP as a combined feature vector.In addition,while combining various features,a consideration has also been given about how to adjust their dimensional effect.