It is found that the detection using basic spectral entropy becomes difficult and inaccurate when speech signals are contaminated by high noise. This paper presents an improved entropy-based algorithm. The way to compute spectral probability density function of entropy is altered by the introduction of a positive constant. The modification improves the discriminability between speech and noise and the robustness of entropy so that it becomes easier to set thresholds. Experiment results reveal the validity of the improved entropy and prove that the improved entropy outperforms basic entropy. Moreover, the improvement of accurate rate (5db SNR) reaches 12.9% for the detection of start and end points averagely comparing with a pure energy-based algorithm.