The great phonetic variability in functionally identical phonological units is one of the major problems in automatic speech recognition. Why isn't it also a problem when humans hear speech? The answer could be that human listeners somehow normalize the variable speech signal. We believe that many sound changes in the history of languages give evidence of listeners attempting to normalize heard speech, the process making itself apparent when it goes awry. We believe that dissimilation - removal of a feature from one of two sites in a word where it was distinctive - is due to the listener mistakenly attributing the feature's presence in one site as non-distinctive spill-over from the other site. We report attempts to duplicate two dissimilatory sound changes via speech perception experiments.