Better Retrieval, Worse Robustness:How Multi-hop RAG Amplifies Upstream ASR Errors

This study evaluates whether multi-hop retrieval-augmented generation becomes more robust or more fragile when spoken queries contain automatic speech-recognition errors. Its controlled suite contains 12,000 spoken queries from HotpotQA, 2WikiMultiHopQA, and MuSiQue, synthesized in four English accents with Microsoft Edge TTS, transcribed by Whisper-large-v3, and answered by OpenAI’s gpt-4o-mini. The systems compare Naive RAG with HippoRAG2’s entity graph, IRCoT’s iterative reformulation, and their combination. Higher word error rate consistently causes larger F1 losses: on Nigerian-accented input, the combined IRCoT+HippoRAG2 system loses 0.142 on HotpotQA, 0.195 on 2WikiMultiHopQA, and 0.072 on MuSiQue relative to clean text, making its degradation 36.5%, 42.3%, and 67.4% larger than Naive RAG’s. Although structurally richer systems retain higher absolute F1, they amplify upstream errors because corrupted query entities disrupt graph linking and successive reformulations; entity corruption appears in at least 87% of 2WikiMultiHopQA degradation cases. N-best decoding and Double Metaphone-based phonetic entity correction recover no more than 12% of the clean-to-ASR gap. Validation on 500 real Nigerian utterances found entity mistranscription in 51.8% of utterances containing named entities, while replacing Whisper with Meta’s SeamlessM4T-v2-large preserved the same robustness ranking. The central implication is that better retrieval architecture can produce worse error robustness in voice-driven RAG.