Abstract: In view of the estimation problem of a priori signal to noise ratio parameter in noisy speech enhancement, by combining two step noise reduction technique and the Gaussian statistical model, a novel estimation algorithm for a priori signal to noise ratio was proposed in frequency domain. Based on the estimation theory of minimum mean square error, the presented algorithm computes directly the square of the clean speech component in its second step to refine the estimated a priori signal to noise ratio of the decision directed approach, and thus the drawback of the two step noise reduction method whose performance is restricted on the gain function is removed. Moreover, the proposed algorithm has good performance in highly reducing the music noise of the output speech while the advantages in noise suppression are kept. Experimental results under different kinds of noisy conditions demonstrate that the performance of the proposed method outperforms that of the classic decision directed method and the recently proposed two step noise reduction technique in both objective and subjective tests.