One of the central problems in speech perception concerns variability, specifically, the mapping relations between acoustic attributes of the signal and discrete linguistic categories resulting from perceptual analysis. Previous accounts of speech perception have treated variability as an undesirable source of noise that must be reduced or eliminated through perceptual normalization. In this paper, I present an alternative view of the categorization process in speech perception that is based on findings obtained from non-analytic approaches to cognition. According to this view, variability in speech is considered to be "lawful" and "informative" for perceptual analysis. Perceptual normalization, as discussed in past theoretical accounts, may not involve a true "loss" of information but rather may entail the encoding of specific instances and the details of perceptual analysis. This approach to categorization in speech perception provides a new way of dealing with a number of long-standing problems in the field.