The invention discloses an endpoint detecting method for voice recognition, which comprises the steps of: collecting background noises and a voice signal containing noises; analyzing the features of the background noises and the voice signal containing noises; extracting parameters of the linear predictive model of the background noises or the LPC (linear predictive coding) coefficient, namely the linear predictive coding coefficient, as the linear predictive template of the background noise; and determining the endpoint of the voice signal containing noises. In the method, the linear predictive coefficient of each frame of voice containing noises is compared with the parameter of the background noise template, and the compared result is treated as a feature value. When the variation of the feature value exceeds a set value, the mark which means that the voice endpoint is detected can be further taken as the background noise template according to the variation of the background noise, namely the linear predictive model which corrects the background noise. According to the invention, the endpoint of the speech voice signal of people can be well detected in the environment with the background noises.