Speakers usually produce words in connected speech. In such contexts, the form in which many words are uttered is influenced by the phonological properties of neighboring words. The current article examines the representations and processes underlying the production of phonologically constrained word form variations. For this purpose, we consider determiners whose form is sensitive to phonological context (e.g., in English: "a car" vs. "an animal"; in French: "le chien" "the dog" vs. "l'âne" "the donkey"). Two hypotheses have been proposed regarding how these words are processed. Determiners either are thought to have different representations for each of their surface forms, or they are thought to have only 1 representation while other forms are generated online after selection through a rule-based process. We tested the predictions derived from these 2 views in 3 picture naming experiments. Participants named pictures using determiner-adjective-noun phrases (e.g., "la nouvelle table" "the new table"). Phonologically consistent or inconsistent conditions were contrasted, based on the phonological onsets of the adjective and the noun. Results revealed shorter naming latencies for consistent than for inconsistent sequences (i.e., a phonological consistency effect) for all the determiner types tested. Our interpretation of these findings converges on the assumption that determiners with varying surface forms are represented in memory with multiple phonological-lexical representations. This conclusion is discussed in relation to models of determiner processing and models of lexical variability.