We discuss three stacking deconvolution techniques for enhancing signal of Earth structure contained in the seismograms of teleseismic body waves recorded by a seismometer array. The goal of the deconvolution is to remove from the seismograms the complications due to the source. For three-component data, the removal of the source can be efficiently achieved by deconvolving one component from another, as in the receiver function analysis. In the case of an array, if the structure beneath individual stations varies, the source wavelet can be estimated and removed from the seismograms by stacking. This allows structural analysis in the situations when only single-component seismometers are used or when a particular component can provide additional independent information. Here we compare the efficacy of conventional time-domain stacking, cepstral averaging, and autocorrelation stacking. The latter is a new approach developed here. For cepstral averaging, we present conditions that stabilize the convergence. Examples are presented using synthetics and real data from a broadband seismic array experiment conducted in the Cascadia subduction zone (Oregon) in 1993 to 1994. Data from subduction zones are particularly well suited because of the spatially rapidly varying structure. We find that all three methods work well for low frequencies, which are not sensitive to the time alignment of the seismic traces. The latter two methods do not require precisely aligned seismograms, and their advantage becomes evident at higher frequencies where the appropriate time alignment is harder to achieve. At high frequencies, time-domain stacking is likely to fail.