The mapping of acoustic speech signals onto linguistic representations is highly complex, and a number of competing models of both system architecture and auditory processing mechanisms have been proposed. Detailed investigation of processing in both normal and aphasic listeners provides insights into the nature of the underlying processing mechanisms, and can be used to test the empirical validity of theoretical models. Three related experiments were designed to explore aspects of the complex interrelations between phonetic, lexical and semantic levels of representation in the auditory speech processing of five adults with chronic aphasia and ten controls. The first experiment explored whether the effects of word frequency and imageability that affect recognition of words heard in isolation exert the same influence when words are heard in the context of a meaningful sentence. The second experiment compared discrimination of voice, place and manner contrasts in nonword and word minimal pairs, to explore the effect of the lexical status of the carrier syllable on phonological encoding. The third experiment used a picture-word verification task to explore the effects of semantic contexts provided by pictures on discrimination. Stimuli were closely matched across experiments two and three to allow detailed comparison of lexical and semantic influences on processing. Accuracy and reaction time data were collected, with control group data indicating normal patterns of performance. Aphasic data were analysed mainly as a series of single cases, and interpreted in the light of each individual's performance on a range of language processing assessments. The results revealed a number of different patterns in the effects of linguistic context for control and aphasic listeners, with aphasic listeners showing greater influences of lexical and semantic contexts on processing. The results of all three experiments are discussed in relation to competing theoretical models of auditory processing. It is argued that a distributed connectionist model currently accounts for the data more effectively than either localist connectionist or cognitive neuropsychological models.