Theoretical phonology has at its heart the assumption of two separate levels of speech sound representations: one less abstract, phonetic, 'surface' level, and one more abstract, phonological, 'underlying' level. Descriptions of phonological neutralisation processes such as German final devoicing hinge on the mappings between these two levels, as a single phonetic realisation can be mapped onto more than one underlying representation. Neurobiological investigations of speech sound representations have unearthed apparently selective, categorical neural responses that are held to reflect phonological processing. This study tests this conclusion by examining native German listeners' neural responses to words beginning with voiced /d/ versus voiceless /t/, where there is an acoustic as well as phonological distinction, and comparing these to responses to words ending with /d/ versus /t/, where the acoustic contrast is neutralised. A difference in brain response to these two final sounds, in the absence of acoustic differences, would be consistent with the involvement of abstract phonological representations in speech perception.Analysis of the MEG data using decoding with a binary classifier and temporal clustering does indeed reveal differences in neural response to words containing neutralised final stops as well as to non-neutralised initial stops. In practice, however, it is difficult to separate out the contribution of abstract phonological representations from responses to the acoustics of the rest of the word and processes involved in lexical access. In particular, the strongest predictor of decoding accuracy for the final /t/ versus final /d/ contrast seems to be the acoustics of the penultimate sound, as much as the identity of the final phoneme. These results indicate that neutralisation, far from being a purely phonological phenomenon, is in fact the result of, and a contributor to, the brain's simultaneous accessing of multiple levels of linguistic representation during speech perception, facilitated by predictive processing and statistical regularities in the speech input. We predict that neutralisation is most likely to occur not only when acoustic cues to a contrast are suboptimal, but also when the two sounds involved tend to co-occur with very different sounds in their immediate phonetic environment. This greater phonotactic predictability means that preserving the acoustic contrast is less crucial for word recognition. [The dissertation citations contained here are published with the permission of ProQuest LLC. Further reproduction is prohibited without permission. Copies of dissertations may be obtained by Telephone (800) 1-800-521-0600. Web page: http://bibliotheek.ehb.be:2222/en-US/products/dissertations/individuals.shtml.]