Since the widely used Hidden Markov Model（HMM）in speech recognition is first order Markov Model

it can not fully model the temporal dependence of speech signal. Although HMM can be extended to higher order Markov Model theoretically

the exponential increase of required computation and memory makes it difficult to use.Therefore

a new model is proposed in this paper

it is constructed by combining HMM with a multi-variable Gaussian density which can depict the temporal dependence of speech signal. The reasonableness of the new model is discussed theoreically. The experiment for all Chinese syllables with tone disregarded（total of 409 syllables）recognition shows that recognition rate of the new model is always significantly better than that of HMM.Furthermore

a discrete smoothed statisical histogram is used to model the state duration

because we found in the experiment that continuous density function

such as Gaussian

Gamma etc.

can not satisfactorily depict the state duration of either HMM or the new model.