Abstract
Single-channel speech dereverberation has been a challenging problem. In this letter, we describe a frequency-domain algorithm which efficiently estimates an acoustic reverberation filter to dereverberate speech by using its inverse filter. Since a reverberation filter is generally much shorter than the corresponding dereverberation filter, the algorithm may converge into a more exact filter and remove reverberant components better than the conventional methods. Simulation on noisy reverberant speech illustrates that the described method can improve speech dereverberation performance in comparison with the method which estimates a dereverberation filter.