The patterns of semantic errors in speaking and writing are used to constrain claims about the structure of lexical access mechanisms in speech and written language production. It is argued that it is not necessary to postulate a modality-neutral level of lexical representation (lemma) that is intermediate between lexical-semantic representations and modality-specific lexical representations. A dual-stage access model is proposed in which the first stage involves the selection of semantically and syntactically specified, modality-specific lexical forms, and the second stage involves the selection of specific phonological (orthographic) content for the selected lexemes.