The properties of linguistic attention were examined by recording event-related brain potentials (ERPs) to probe stimuli mixed with dichotically presented prose passages. Subjects either shadowed (repeated phrase by phrase) or selectively listened to one passage while ERPs were recorded from electrodes overlying midline sites, left-hemisphere speech areas, and corresponding areas of the right hemisphere. Mixed with each voice (a male voice in one ear, a female voice in the other) were four probe stimuli: digitized speech sounds (but or /a/ as in father) produced by the same speaker and tone bursts at the mean fundamental and second formant frequencies of that voice. The ERPs elicited by the speech probes in the attended ear showed an enhanced negativity, with an onset at 50 ms-100 ms and lasting up to 800 ms-1,000 ms, whereas the ERPs to the second formant probes showed an enhanced positivity in the 200 ms-300 ms latency range. These effects were comparable for shadowing and selective listening conditions and remained stable over the course of the experiment. The attention-related negativity to the consonant-vowel-consonant probe (but) was most prominent over the left hemisphere; other probes produced no significant asymmetries. The results indicate that stimulus selection during linguistic attention is specifically tuned to speech sounds rather than simply to constituent pure-tone frequencies or ear of entry. Furthermore, it appears that both attentional set and stimulus characteristics can influence the hemispheric utilization of stimuli.