This paper describes a large computer-coded conversational speech data base and results of testing phonological rules on that data base. The study shows that frequency of rule application depends not only on phonological environments, but also on frequency of occurrence of specific words. That is, some rules are highly word-dependent, others are more phonologically governed. This relationship between rule application and word frequency suggests that different kinds of phonological variation should be represented in different ways in the lexicon and phonological rule components of a speech understanding system. (Author)