Abstract: The search on the minimum of the speech power spectrum involved in noise is the foundation of noise estimation. In order to improve the estimation accuracy of the non stationary noise and reduce the delay of noise estimation in the case of noise bursts, this paper proposes a minimum search method consisting of two parallel searches with different window lengths. The results of the minimum search are determined by the ones of the two parallel searches and the binary decision about speech presence based on the noise classification. It is shown from experimental results that for the highly non stationary noise, the proposed method can effectively reduce the delay of noise estimation and improve the quality of enhanced speech significantly.