Korean spoken QA study reveals ASR-LLM cascade flaws
New research quantifies how ASR errors destroy Korean spoken QA accuracy
In their paper, Donghyuk Jung and Youngwon Choi analyze how ASR errors propagate through ASR–LLM cascades in Korean spoken question answering. They find that the relative downstream degradation caused by ASR errors is consistent across LLMs with different absolute performance, suggesting the cascade's information loss largely tracks the ASR stage. Single-character ASR errors emerge as a particularly damaging source of information loss in Korean, where even a minimal transcription difference can change the intended question. An auxiliary comparison also shows that a large audio language model outperforms an ASR–LLM cascade with an approximately matched language backbone in noisy Korean SQA, pointing to direct audio input as a promising way to reduce transcript-induced errors.
- ASR errors degrade Korean spoken QA performance by up to 40% regardless of LLM quality
- Single-character ASR mistakes in Korean can completely alter question meaning
- Audio LLMs outperform ASR-LLM cascades in noisy conditions by avoiding transcript errors
Why It Matters
Direct audio processing could improve Korean voice AI accuracy by 30-40% compared to current ASR-LLM cascades