Two experiments critically examined a predictive-coding based account of the vulnerability of short-term memory (STM) to auditory distraction, particularly the disruptive effect of changing-state sound on verbal serial recall. Experiment 1 showed that providing participants with the opportunity to predict the contents of an imminent spoken distractor sentence via a forewarning reduced its particularly disruptive effect but only to the same level of disruption as that produced by simpler changing-state sequences (a sequence of letter-names). Moreover, a postcategorically unpredictable changing-state sequence (e.g., “F, B, H, E, . . .”) was no more disruptive than a postcategorically predictable sequence (“A, B, C, D, . . .”). Experiment 2 showed that a sentence distractor was disruptive regardless of whether participants reported adopting a serial rehearsal strategy to perform the focal task (in this case, a missing-item task) whereas, critically, the disruptive effect of simpler changing-state sequences was only found in participants who reported using a serial rehearsal strategy. Moreover, when serial rehearsal was not used to perform the focal task, the disruptive effect of sentences was completely abolished by a forewarning. These results indicate that predictability plays no role in the classical changing-state irrelevant sound effect and that foreknowledge selectively attenuates a functionally distinct stimulus-specific attentional-diversion effect. As such, the results are at odds with a unitary, attentional account of auditory distraction in STM and instead strongly support a duplex-mechanism account.