It has been argued that partial information from a lexical- semantic route and partial phonological information from sublexical routines can combine to facilitate lexical access in word production (e.g. Hillis & Caramazza, 1995; Nickels, 1992). Hillis and Caramazza (1991) suggested that in word reading or word repetition, if partial semantic activation leads to semantic competitors being equally activated at the level of the lexicon then the provision of at least some phonological information would usually avert a semantic error occurring. It follows then that semantic errors are only predicted to occur if nonword reading or repetition are abolished. This paper describes a young man, MI, who following brain damage produces semantic errors in the repetition of single words deep dysphasia. This is ascribed to two specific deficits within a language processing model: one in accessing semantics from auditory input despite successful retrieval of the lexical representations and a second in the repetition of nonwords. However, although MI’S nonword repetition is impaired he can repeat some nonwords and errors are generally phonologically related to their targets. It is argued that these data are incompatible with the theory proposed by Hillis and Caramazza (1991).