Three experiments investigated the use of lexical stress in auditory word recognition. Speech voicing continua were created in which lexical status resulted in one end-point constituting a real word and the other end a nonword (e.g. diGRESS-tiGRESS, in which digress is a real word, and DIgress-Tigress, in which tigress is a real word). Subjects' identification of the initial segment of these items was biased in the midrange of the continua in that they were more likely to report a segment that resulted in a real word than one that resulted in a nonword. Alternative acoustically based explanations are discounted in favor of a lexically based account of the data. Possible mechanisms underlying the effect of lexical stress on speech perception are discussed.