Systemic research has been done on the endpoint detection algorithms of speech in the noise environment. traditional method of short-time energy and short-time zero-crossing rate of double threshold can not detect speech signal endpoint in low SNR. To solve this problem, A improved algorithm of spectral entropy-double threshold is proposed in this paper, the diagram of realization of this algorithm is presented and simulated in MATLAB. Simulation results show that the algorithm perform well on anti-noise,they can still accurately distinguish between voice and noise, so the effectiveness and high efficiency of the improved algorithm is proved.