A proposed model of the process by which a human listener identifies the sequence of words in running speech is described. It is assumed that words are stored in memory as sequences of bundles of binary distinctive features. In a given language these features are a subset of a universal inventory of features. These features are of two kinds: articulator-free and articulator-bound. Identification of words is achieved by extracting cues from the signal, identifying the features, and matching against the lexicon. The basic or defining acoustic correlates of articulator-free features are acoustic landmarks in the signal. Acoustic correlates of articulator-bound features for a segment are generally found in the vicinity of these landmarks. In running speech there is some variability in the acoustic manifestation of the features, particularly articulator-bound features. This variability arises for two reasons: (1) Although one can postulate "defining" acoustic and articulatory attributes for each feature, these attributes are often supplemented by the introduction of additional gestures that enhance the perceptual saliency of the feature, and this enhancement may be context dependent. (2) The gestures used to produce a sequence of segments may overlap, causing some of the cues for the features to be weakened or obliterated. Examples of feature-defining articulatory and acoustic attributes, enhancement gestures, and the effects of gestural overlap are given. Knowledge of these sources of variability can help to guide the selection and weighting of acoustic cues.