Speech error data and empirical studies indicate that the scope of planning is larger for semantic/syntactic than phonological form representations in speech production. R. Martin and colleagues have demonstrated that some patients show dissociable impairments in the retention of semantic and phonological codes. The effect of these STM deficits on speech production was investigated using phrase and sentence production paradigms which manipulated semantic and phonological factors of the words to be produced. The first four experiments used a phrase production paradigm in which subjects named pairs of pictures and words manipulated for categorical relatedness (Experiments 1 and 2) or frequency (Experiments 3 and 4). The findings suggested that subjects planned both lemmas in these phrases but only the initial lexeme. Planning two lemmas created difficulty (large categorical interference effects) for patients with semantic retention deficits but not for patients with phonological retention deficits, supporting the contention that the semantic retention buffer involved in word span tasks is involved in speech production planning. In Experiment 5, subjects produced sentences having the content words in separate phrases to describe the configuration of two pictures. Control subjects and patients with phonological retention deficits did not show interference from categorically related pictures suggesting that they did not plan the final lemma of the sentence before articulation. ML still showed a large interference effect which reversed to facilitation when the pictures remained on the screen during his response. It was suggested that ML is forced to plan multiple lemmas before the pictures disappear, but when the pictures remain in view he may plan in a phrase-by-phrase fashion. In Experiment 6, ML showed a large benefit in onset latency but not accuracy when he could describe pictures in a sentence compared to phrase format whereas EA showed no such benefit. SJD also showed a benefit in accuracy for sentences compared to phrases, but her errors were qualitatively different than those of ML. The findings suggest that semantic retention is critical for speech planning for constructions requiring that multiple lemmas be planned in advance, while phonological retention is less critical because of the smaller scope of planning at this level.