Reverberation can severely degrade speech signals captured by microphones in a room. A popular dereverberation algorithm is the weighted prediction error (WPE) algorithm. When considering spatially distributed microphones, a similar dereverberation performance may be achieved using WPE with only a subset of all available microphones. In this paper, we propose a WPE microphone subset selection method using a novel block-swap subset selection algorithm. The proposed method performs a local search using an efficient computation of the WPE microphone subset selection cost. Experimental results using noisy reverberant speech recorded by 18 spatially distributed microphones show that the proposed microphone subset selection method yields frequency-independent and frequency-dependent subsets with a similar dereverberation performance to that of the optimal subsets computed using exhaustive search.