On discussing the defects of the traditional voice activity detection method based on cepstrum distance, this paper proposes an improved project. In this method, an accurate SNR estimation is made at first, and the discriminative threshold is fixed according to the relationship between SNR and the threshold, then the accurate end-point can be confirmed, in the end the simulation experiments are made with a lot of noisy speech signals ranging from - 5dB to 20 dB in three representative noisy environments. The result shows that the comparatively accurate speech endpoints can be detected by the above method in different SNR environments.