Outlined in this paper is a novel approach to speech dereverberation when an estimate of the source-receiver transfer function is known. It is a two-stage algorithm based on the minimum phase/allpass decomposition of a mixed phase room impulse response (RIR). The reverberant speech is first filtered with the inverse minimum phase component of the RIR. Then a Non-Negative Matrix Factorization (NMF) based denoising approach is used to remove artifacts associated with the allpass component of the RIR from the inverse filtered speech. This approach was tested on speech convolved with synthetically generated room impulse responses. The results of these tests were analyzed using objective measures and listening tests both of which indicate that this approach leads to significant enhancement of the reverberant speech. ©2008 IEEE.