RCbench: Benchmarking Retrospective Clarification in ASR

Human speakers often clarify newly introduced or ambiguous words later in an utterance, for example by providing an explanation, an acronym expansion, or a spelling cue. However, current ASR systems are generally unable to incorporate such later clarifications to revise earlier parts of the transcript. We define this problem as retrospective clarification, and present RCBench, a benchmark consisting of three scenarios: acronym disambiguation, disambiguation by semantic explanation, and disambiguation by spelling. On RCBench, we find a substantial gap between current ASR systems and human participants. In particular, within the same utterance, ASR baselines achieve 60–90% accuracy on control words that do not require retrospective clarification, suggesting that the main challenge lies in revising specific target words based on later clarification. Source code: https://anonymous.4open.science/r/RCbench