In this work we have try to use the time dependant linear prediction technique in order to automatically recognize complete isolated words. We have selected the autocorrelation method because it works faster than convariance since it gives a correlation matrix with a high redundancy. We consider a word as a point in a 72-dimensions space, where each dimension corresponds to one coefficient of the time varying linear prediction. In order to obtain the reference patterns we have try two methods. The first one is selecting among the versions of a word the one with minimal distance added among versions. The second method consists on computing the gravity center of the several versions of a word. Between these two methods the better has been the second one. As far as the base functions is concerned we have used potential functions, trigonometric functions, Walsh functions and Haar functions.