Listening to a speech message requires the accurate selection of the relevant auditory input especially when distracting background noise or other speech messages are present. To investigate such auditory selection processes we presented three different speech messages simultaneously spoken by different actors at separate spatial locations (-70, 0, 70/ azimuth). Stimuli were recorded using an artificial head with microphones embedded in the "auditory canals" to capture the interaural time and level differences as well as some of the filter properties of the outer ear structures as auditory spatial cues, thus creating a realistic virtual auditory space. In a given experimental run young healthy participants listened via headphones and either attended to the rightmost or the leftmost message in order to comprehend the story. Superimposed on the speech messages task irrelevant probe stimuli (syllables sharing spatial and spectral characteristics, 4 probes/s) were presented that were used for the generation of event-related brain potentials computed from 29 channels of EEG. ERPs to probe stimuli were characterized by a negativity starting at 250 ms with a contralateral frontal maximum for the probes sharing spatial/spectral features of the attended story relative to those for the unattended message. The relatively late onset of this attention effect was interpreted to reflect the task demands in this complex auditory environment. This study demonstrates the feasibility to use virtual auditory environments in conjunction with the probe technique to study auditory selection under realistic conditions.