These studies investigated the effects of several sources of naturally occurring variability in speech, both in isolation and in combination, on the perceptual identification of spoken words. Identification accuracy was poorer for word lists containing tokens produced by multiple talkers or at multiple rates compared to the corresponding single-rate/single-talker conditions. Furthermore, simultaneous variations along both rate and talker produced greater reductions in perceptual identification than either source alone. In contrast, variability due to overall amplitude did not significantly alter subjects' ability to correctly identify stimulus items. These findings suggest that the acoustic waveform is subjected to one or more transformations that act upon item-specific information in the signal prior to arriving at phonetic decisions. Implications of the results for models of speech perception are discussed.