On the basis of the dereverberation method based on TF-GSC and minimum phase decomposition, considering the insensitivity of the human ear to the speech phase information, an improved method that Only use the phase of a reverberant speech to replace the phase information of the whole component is proposed. The time domain waveforms produced by all the dereverberation method had a comprehensive comparison. At the same time all the method is evaluated by simulation using two subjective and three objective dereverberation evaluation indices. Results show that the improved method has better performance of dereverberation and save about half of the computational complexity so that the real-time processing speed of the algorithm is greatly improved.