Methods are proposed for incorporating phonetic and phonological knowledge into models for lexical access. These include representing the lexicon in terms of linguistically-motivated segments and features, so that dialectal and context-conditioned modifications of features can be accounted for in a direct way. Acoustic analysis to retrieve the segments and features from an utterance requires the detection of landmarks in the signal, followed by measurement of several cues for each feature. These cues are combined to produce an integrated property that determines the feature that is implemented by the talker. Examples of the required acoustic cues are given. Comparison is made with current approaches to lexical access based on statistically-motivated models.