Abstract: The performances of the traditional speech endpoint detection algorithms will decline sharply in lower SNR environments. This paper proposes two new methods: feature weighted likelihood and divergence based dimension reduction to improve detecting performance in noise. The weighted likelihood method can increase the proportion of dynamic feature in likelihood score, consequently improves noise robustness of endpoint detection. Reducing these feature dimensions with smaller divergence will degrade the performance little, but can decrease the computation a lot. Weighted likelihood is also effective for reduced feature. The experimental results on Aurora2 show feature weighted likelihood can remarkably improve the detection performance when the model trained on clean data is used to detect speech endpoint in noise. The performance using weighted likelihood on reduced feature is comparable to that on original feature. This proves the method is robust for noise.