The present study examined the extent to which word production and recognition rely on shared representations in lexical access by examining cross-modality transfer effects and frequency effects in a training paradigm. Participants were trained in reading high- and low-frequency words in a lexical decision task and were subsequently tested in producing picture names and vice versa, both in their second (Experiment 1) and in their first lan- guage (Experiment 2). The same pattern of results was found for first and second language processing. Both tasks showed strong, within-modality repetition effects with faster responses and smaller frequency effects for repeated items. Training with repeated lexical decision, sped responses, and reduced the size of the frequency effects in subsequent picture naming. In contrast, training with repeated picture naming sped responses in lexical decision, but did not significantly decrease frequency effects. The results imply an amodal representation (lemma) that is shared between production and recognition and is not sensitive to word frequency. Also, they imply that a frequency sensitive phonological representation (lexeme) is activated automatically during visual word recognition.