Endpoint detection based on LPC distance will first frame the signal processing, and then according to the LPC distance set threshold to determine whether each frame is the endpoint of the voice. However, it will have the following questions in the signal frame: if a larger window length is used, the total number of frames will be reduced, but the error in the endpoint detection will also increase, especially for the detection of the origin of the speech. If the frame length selection is too small, the calculation amount and the recognition speed are increased. Aiming at this problem, this paper proposes an improved endpoint detection algorithm combining the variable window length idea with the LPC distance endpoint detection algorithm. The method does not use the uniform window length in the frame, but in the quiet section uses a large window length quickly approaching the transition section. Then after the transition ,the small window length is used slowly to look for the starting point of the voice signal, the other stage is the use of conventional windows long. Simulation results show that the improved endpoint detection method proposed in this paper is superior to the other two endpoint detection methods.