This paper presents a novel post-filter for noise reduction. A subspace based noise estimation method is developed with the use of multiple statistical distributions to model the speech and noise. The signal-plus-noise subspace dimension is determined by maximizing the target speech presence probability in noisy frames, so as to estimate the noise power spectrum for post-filter design. Then, masking property is incorporated in the post-filter technique for residual noise shaping. Experimental results show that the proposed scheme outperforms the baseline systems in terms of various quality measurements of the enhanced speech.