A feature-based model for accessing words in a lexicon from measurements on the speech signal is proposed, and the implementation of some components of the model is described. Words are represented in a lexicon in terms of hierarchies of distinctive features arranged in matrix form. Acoustic measurements are made on an utterance in order to estimate the pattern of features implemented by the speaker, and access to lexical items is based on these hypothesized features. The acoustic measurements are of two kinds: (1) they identify and classify critical locations in the signal, and (2) they indicate the activity of articulatory structures in the vicinity of these critical locations. Proposed procedures for representing the lexicon and for accessing words from the lexicon have the potential of being relatively insensitive to speaking style and to context.