Speech interfaces put ASR in front of retrieval, so the query a RAG system sees is already corrupted. We test whether the standard multi-hop methods, entity-graph linking and iterative reformulation, absorb that corruption or amplify it.</p>\n<p>The findings suggest that they amplify it. The structurally richer configurations generally reach higher absolute F1 under ASR input, but they give up more of their clean-text advantage: the gap from clean text to the highest-WER accent is 36-67% larger under IRCoT+HippoRAG2 than under naive dense retrieval, on all three benchmarks. Corruption of one or more query entities is the dominant failure mode, accounting for 87-96% of degradation cases on 2WikiMultiHopQA across all four methods.</p>\n<p>Two lightweight surface-form mitigations, N-best decoding and phonetic entity correction, leave most of the gap intact, which points at retrieval structure rather than transcription surface form.</p>\n<p>We release the code and a transcript dataset covering 3 multi-hop QA benchmarks x 4 English accents (3,000 questions, 12,000 rows).</p>\n","updatedAt":"2026-08-25T05:19:29.050Z","author":{"_id":"679caa50208bfa0539ca05c8","avatarUrl":"/avatars/5e9d76169dbcd8c414d8cddaa2029c0a.svg","fullname":"Zhenghua Bao","name":"KingZ23","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":1,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8864525556564331},"editors":["KingZ23"],"editorAvatarUrls":["/avatars/5e9d76169dbcd8c414d8cddaa2029c0a.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2608.22872","authors":[{"_id":"6a8d09365add2537c32e96e6","user":{"_id":"679caa50208bfa0539ca05c8","avatarUrl":"/avatars/5e9d76169dbcd8c414d8cddaa2029c0a.svg","isPro":false,"fullname":"Zhenghua Bao","user":"KingZ23","type":"user","name":"KingZ23"},"name":"Zhenghua Bao","status":"claimed_verified","statusLastChangedAt":"2026-08-25T08:13:56.717Z","hidden":false}],"mediaUrls":["https://cdn-uploads.huggingface.co/production/uploads/679caa50208bfa0539ca05c8/z-Aq63JmAHk8PRuEn6tTV.png"],"publishedAt":"2026-08-24T00:00:00.000Z","submittedOnDailyAt":"2026-08-25T00:00:00.000Z","title":"Better Retrieval, Worse Robustness:How Multi-hop RAG Amplifies Upstream ASR Errors","submittedOnDailyBy":{"_id":"679caa50208bfa0539ca05c8","avatarUrl":"/avatars/5e9d76169dbcd8c414d8cddaa2029c0a.svg","isPro":false,"fullname":"Zhenghua Bao","user":"KingZ23","type":"user","name":"KingZ23"},"summary":"Speech-based applications pass spoken queries through automatic speech recognition (ASR) before any retrieval module, so ASR errors enter the pipeline as a fixed upstream constraint. We empirically test whether two extensions to standard retrieval-augmented generation (RAG), entity-graph linking and iterative reformulation, absorb or amplify these errors. Using four English accents synthesized through neural TTS, we evaluate four RAG configurations on three multi-hop QA benchmarks (HotpotQA, 2WikiMultiHopQA and MuSiQue) against a clean-text oracle. Although the structurally richer configurations generally retain higher absolute F1 under ASR input, both extensions amplify the error: the F1 gap from clean text to the highest-WER accent is 36-67% larger under their combination than under naive dense retrieval, on all three benchmarks. The dominant failure mode is corruption of one or more query entities, accounting for 87-96% of degradation cases on 2WikiMultiHopQA across all four methods. Two lightweight surface-form mitigations leave most of the gap intact, indicating that downstream retrieval structure amplifies remaining entity errors. We release code and data at https://github.com/ZhenghuaBao/spoken-multihop-rag .","upvotes":3,"discussionId":"6a8d09365add2537c32e96e7","projectPage":"https://huggingface.co/datasets/KingZ23/spoken-multihop-rag","githubRepo":"https://github.com/ZhenghuaBao/spoken-multihop-rag","githubRepoAddedBy":"user","ai_summary":"Retrieval-augmented generation extensions amplify automatic speech recognition errors in spoken multi-hop question answering, primarily through corrupted query entities.","ai_keywords":["retrieval-augmented generation","entity-graph linking","iterative reformulation","automatic speech recognition","multi-hop QA","dense retrieval","neural TTS"],"ai_summary_model":"thinkingmachines/Inkling-Small","githubStars":1},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"679caa50208bfa0539ca05c8","avatarUrl":"/avatars/5e9d76169dbcd8c414d8cddaa2029c0a.svg","isPro":false,"fullname":"Zhenghua Bao","user":"KingZ23","type":"user"},{"_id":"63ac5701c21e60a3e9b58aa7","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/63ac5701c21e60a3e9b58aa7/g6EX7diOpuA94R2ab-rZC.png","isPro":true,"fullname":"Dipankar Sarkar","user":"dipankarsarkar","type":"user"},{"_id":"65224256377bffdc599901c9","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/65224256377bffdc599901c9/xUM64_ugWwt4NO1F6DOcc.jpeg","isPro":false,"fullname":"chentanyu","user":"sailorjs0804","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2608/2608.22872.md","query":{}}">
Better Retrieval, Worse Robustness:How Multi-hop RAG Amplifies Upstream ASR Errors
Abstract
Retrieval-augmented generation extensions amplify automatic speech recognition errors in spoken multi-hop question answering, primarily through corrupted query entities.
Speech-based applications pass spoken queries through automatic speech recognition (ASR) before any retrieval module, so ASR errors enter the pipeline as a fixed upstream constraint. We empirically test whether two extensions to standard retrieval-augmented generation (RAG), entity-graph linking and iterative reformulation, absorb or amplify these errors. Using four English accents synthesized through neural TTS, we evaluate four RAG configurations on three multi-hop QA benchmarks (HotpotQA, 2WikiMultiHopQA and MuSiQue) against a clean-text oracle. Although the structurally richer configurations generally retain higher absolute F1 under ASR input, both extensions amplify the error: the F1 gap from clean text to the highest-WER accent is 36-67% larger under their combination than under naive dense retrieval, on all three benchmarks. The dominant failure mode is corruption of one or more query entities, accounting for 87-96% of degradation cases on 2WikiMultiHopQA across all four methods. Two lightweight surface-form mitigations leave most of the gap intact, indicating that downstream retrieval structure amplifies remaining entity errors. We release code and data at https://github.com/ZhenghuaBao/spoken-multihop-rag .
Community
Speech interfaces put ASR in front of retrieval, so the query a RAG system sees is already corrupted. We test whether the standard multi-hop methods, entity-graph linking and iterative reformulation, absorb that corruption or amplify it.
The findings suggest that they amplify it. The structurally richer configurations generally reach higher absolute F1 under ASR input, but they give up more of their clean-text advantage: the gap from clean text to the highest-WER accent is 36-67% larger under IRCoT+HippoRAG2 than under naive dense retrieval, on all three benchmarks. Corruption of one or more query entities is the dominant failure mode, accounting for 87-96% of degradation cases on 2WikiMultiHopQA across all four methods.
Two lightweight surface-form mitigations, N-best decoding and phonetic entity correction, leave most of the gap intact, which points at retrieval structure rather than transcription surface form.
We release the code and a transcript dataset covering 3 multi-hop QA benchmarks x 4 English accents (3,000 questions, 12,000 rows).
Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.
Tap or paste here to upload images
Cite arxiv.org/abs/2608.22872 in a model README.md to link it from this page.
Cite arxiv.org/abs/2608.22872 in a Space README.md to link it from this page.
Discussion (0)
Sign in to join the discussion. Free account, 30 seconds — email code or GitHub.
Sign in →No comments yet. Sign in and be the first to say something.