Abstract: To improve the accuracy of speech endpoint detection in low signal-to-noise ratio(SNR), a new algorithm was proposed. It combines the short-time energy-zero-product with discrimination information, denoising and endpoint detection simultaneously. Then, the frames judged as noisy ones are rechecked via the discrimination information based on the sub-band energy distribution probability to update the noise energy threshold, thus tracking the change for noise energy more accurate. Simulation results showed that the algorithm proposed is able to detect endpoints accurately even if the SNR changes acutely, and it will serve very well the function of subsequent processing of speech signal.