This paper studied the supervisor flat net for speech recognition, The extension function model,a input pattern time sequence relation model,is discussed. The experiments indicated that the converage performance of our model is superior than the multi-layer perception (MLP).