Analyzing Error Propagation in Korean Spoken QA with ASR-LLM Cascades

The study analyzes error propagation in Korean spoken question answering (SQA) using ASR-LLM cascades, revealing that ASR errors lead to consistent downstream semantic failures, irrespective of the LLM's overall performance. It identifies single-character transcription errors as a significant loss channel in Korean, affecting QA outcomes, and demonstrates that a large audio language model can outperform an ASR-LLM cascade in noisy environments, highlighting the advantages of direct audio input in reducing information loss. This research is crucial for practitioners as it underscores the importance of addressing ASR errors and exploring alternative input methods to improve SQA systems.