The MCRA minimum recursive algorithm is accurate for the noise estimation, and the changes of noise power spectrum in a speech can be tracked accurately. However, if the noise power spectrum increases too much suddenly, the original algorithm needs a period of time to get the accurate noise, and in this adaptive period, it will leave strong residual noise and affect people's hearing experience. This paper introduces a Voice Activity Detection (VAD) algorithm which uses the maximum log-likelihood ratio with energy-zero ratio, and an improved noise estimation algorithm on the basis of MCRA is obtained. Experimental simulation also proves that the improved algorithm is better than the original algorithm in noise estimation speed.