Abstract: In an enclosure room, the speech signal captured by a distant microphone is usually contaminated by reverberation, in which the late reverberation severely degrades the speech intelligibility. Common denoising methods can only remove the additive noise such as white noise, not the reverberation which is the convolution of clean speech and room impulse response. Therefore, dereverberation algorithm should be applied to eliminate the bad effects caused by late reverberations. In this paper, we propose an improved dereverberation algorithm using multiple-step linear prediction. This method modifies the process of pre-whitening, which can raise the first two formants of the speech signal. Experimental results showed that new dereverberation algorithm can remove most of the late reverberations and reserve more low frequency components of clean speech, consequently improve the speech intelligibility.