Event-related potentials (ERPs) were recorded from the scalp to investigate a long-standing controversy in auditory attention research, namely when the 'breakthrough of the unattended' takes place in the human brain. Nine subjects classified visual stimuli appearing 300 ms after task-irrelevant standard tones (80%, i.e. P = 0.8) or novel sounds (20%, i.e. P = 0.2) into odd/even categories. After the recording session, subjects scored the novel sounds as to whether they had any particular meaning (identifiable) or were perceived as a burst of noise (non-identifiable), and performance and ERPs were analysed according to this classification. A control condition, in which the visual stimuli were presented with no sounds, showed that subjects covertly monitored the task-irrelevant sounds during visual task-performance, and a further condition, in which the auditory and visual stimuli appeared regardless of each other, made it possible to trace the processing of the distracters during allocation of attention outside the auditory environment. Results yielded identical N1-enhancement for the two types of novel sounds, indicating similar attention switching triggered to these two types of unexpected sounds. However, there was a stronger orientating of attention towards identifiable novel sounds, as indicated both by behavioural distraction and by larger novelty-P3. Furthermore, this stronger orientating of attention was due to the sounds being contingent on the visual stimuli, as no increase in novelty-P3 to identifiable novel sounds was observed in the control condition, in which the sounds occurred outside the attentional set. Therefore, provided that the N1-enhancement reflects a call for focal attention, and novelty-P3 the effective orientating of attention towards the eliciting sounds, the present results suggest that semantic analysis of significant sounds occurs after a transitory switch of attention towards the eliciting stimuli. Moreover, as the novelty-P3 increase in amplitude was observed only when subjects covertly monitored the sounds, the present data suggest that semantic analysis of irrelevant sounds depends on the top-down cognitive influences of the attentional set.