MedSpeak: A Knowledge Graph-Aided ASR Error Correction Framework for Spoken Medical QA

Amir Rahmani; Chenhan Lyu; Elahe Khatibi; Honghui Xu; Nikil Dutt; Pengfei Zhang; Shiva Shrestha; Yutong Song

arxiv: 2602.00981 · v2 · submitted 2026-02-01 · 💻 cs.CL · cs.AI

MedSpeak: A Knowledge Graph-Aided ASR Error Correction Framework for Spoken Medical QA

Yutong Song , Shiva Shrestha , Chenhan Lyu , Elahe Khatibi , Pengfei Zhang , Honghui Xu , Nikil Dutt , Amir Rahmani This is my paper

Pith reviewed 2026-05-16 09:22 UTC · model grok-4.3

classification 💻 cs.CL cs.AI

keywords ASR error correctionmedical knowledge graphspoken question answeringmedical terminologyLLM reasoning

0 comments

The pith

MedSpeak refines ASR transcripts for medical questions by combining a knowledge graph's semantic and phonetic links with LLM reasoning to raise term accuracy and overall QA performance.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents MedSpeak as a framework that takes noisy automatic speech recognition output from spoken medical questions and corrects it before passing the text to an answer model. It does so by consulting a medical knowledge graph for both meaning connections between terms and sound-alike relationships, then using large language model reasoning to select the most plausible correction. If this works, spoken medical QA systems would make fewer mistakes on domain-specific vocabulary that general ASR engines routinely mangle, improving the reliability of voice-based medical information tools. The authors support the claim with benchmark experiments showing gains in both medical-term recognition and end-to-end answer accuracy.

Core claim

MedSpeak is a knowledge-graph-aided ASR error correction framework that refines noisy transcripts by leveraging semantic relationships and phonetic information encoded in a medical knowledge graph together with the reasoning power of LLMs, thereby improving the accuracy of medical term recognition and overall medical SQA performance.

What carries the argument

The knowledge-graph-guided correction step that pulls both semantic relations and phonetic similarity from the medical graph and feeds candidate fixes to an LLM for final selection.

If this is right

Medical-term recognition accuracy rises on standard spoken medical QA benchmarks.
End-to-end answer prediction performance improves compared with uncorrected ASR output.
The same graph-plus-LLM correction pattern can be reused for other medical spoken-language tasks.
Fewer domain-specific transcription errors reach the answer generation stage.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same correction approach could transfer to spoken QA in other technical domains if suitable knowledge graphs exist.
Real-time deployment would require the graph lookup and LLM call to fit inside acceptable latency bounds for live conversation.
If the knowledge graph is incomplete for rare terms, hybrid methods that also consult external medical lexicons might be needed.

Load-bearing premise

The medical knowledge graph encodes enough accurate semantic relationships and phonetic information to guide effective corrections when combined with LLM reasoning, without introducing new errors.

What would settle it

Run the framework on a medical-speech test set containing terms absent from or poorly connected in the knowledge graph; if term-recognition accuracy and downstream QA scores show no improvement or decline relative to the uncorrected baseline, the central claim is falsified.

Figures

Figures reproduced from arXiv: 2602.00981 by Amir Rahmani, Chenhan Lyu, Elahe Khatibi, Honghui Xu, Nikil Dutt, Pengfei Zhang, Shiva Shrestha, Yutong Song.

read the original abstract

Spoken question-answering (SQA) systems relying on automatic speech recognition (ASR) often struggle with accurately recognizing medical terminology. To this end, we propose MedSpeak, a novel knowledge graph-aided ASR error correction framework that refines noisy transcripts and improves downstream answer prediction by leveraging both semantic relationships and phonetic information encoded in a medical knowledge graph, together with the reasoning power of LLMs. Comprehensive experimental results on benchmarks demonstrate that MedSpeak significantly improves the accuracy of medical term recognition and overall medical SQA performance, establishing MedSpeak as a state-of-the-art solution for medical SQA. The code is available at https://github.com/RainieLLM/MedSpeak.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

MedSpeak combines a medical KG with LLM reasoning to clean up ASR errors in spoken medical QA, but the phonetic correction step looks under-supported and the gains need tighter baselines.

read the letter

MedSpeak proposes a pipeline that pulls semantic relations and phonetic cues from a medical knowledge graph to fix ASR mistakes in spoken medical question answering, then uses an LLM to produce better answers. The core move is practical: treat the KG as an external memory that can catch both meaning and sound-alike errors before the final prediction step. They release code, which makes the work easier to check or extend. That is the main thing worth noting. The framework is a straightforward application of existing tools rather than a new method, but it targets a real pain point in medical speech systems where terminology errors matter. The experiments are said to show gains on benchmarks, and the description of the correction process is clear enough to follow. The main weakness is the phonetic part. Standard medical KGs store synonyms and relations but rarely include pronunciation distances or IPA mappings. If the paper's KG construction does not add explicit phonetic features, the correction reduces to semantic lookup and the LLM can still hallucinate plausible but incorrect terms. The abstract gives no numbers, baselines, or error bars, so the full paper must supply those details and an ablation on the phonetic component to make the claim stick. This is useful for groups already building domain-specific spoken QA tools. It is not reshaping the field, but the concrete setup and code release give it enough substance for a referee to evaluate. I would send it to review with requests for the missing metrics and a clearer account of how phonetic information is actually encoded or approximated.

Referee Report

3 major / 0 minor

Summary. The manuscript proposes MedSpeak, a knowledge graph-aided ASR error correction framework for spoken medical QA. It refines noisy ASR transcripts by leveraging semantic relationships and phonetic information encoded in a medical knowledge graph together with LLM reasoning, with the goal of improving downstream answer prediction. The authors claim that experiments on benchmarks demonstrate significant gains in medical term recognition and overall SQA performance, establishing MedSpeak as state-of-the-art, and release code at https://github.com/RainieLLM/MedSpeak.

Significance. If the empirical claims hold under rigorous evaluation, the work could advance medical spoken QA by providing a practical way to mitigate ASR failures on domain-specific terminology. The combination of structured KG knowledge with LLM reasoning is a timely direction, and the public code release is a clear strength that supports reproducibility and extension.

major comments (3)

[Abstract] Abstract: The central claim that MedSpeak 'significantly improves' medical term recognition and 'establishes' SOTA performance is unsupported by any reported metrics, baselines, error bars, or benchmark details. This absence leaves the primary contribution without visible evidence.
[§3] §3 (KG construction): The framework relies on the medical KG supplying usable phonetic information (e.g., pronunciations, IPA mappings, or phonetic similarity metrics) alongside semantics. Standard resources such as UMLS contain semantic relations and synonyms but rarely encode phonetics; the manuscript must explicitly describe how phonetic features are added or derived, otherwise the correction step reduces to semantic lookup and cannot reliably address sound-alike medical-term errors.
[Experimental Section] Experimental evaluation: The manuscript must include quantitative results (e.g., WER or term-accuracy deltas), comparison tables against relevant baselines, statistical significance tests, and ablation studies on the phonetic component to substantiate the improvement and SOTA assertions.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We have carefully considered each major comment and revised the paper to address the concerns regarding evidence, clarity, and experimental rigor. Our responses are provided point by point below.

read point-by-point responses

Referee: [Abstract] Abstract: The central claim that MedSpeak 'significantly improves' medical term recognition and 'establishes' SOTA performance is unsupported by any reported metrics, baselines, error bars, or benchmark details. This absence leaves the primary contribution without visible evidence.

Authors: We agree that the original abstract lacked specific quantitative support for the claims. In the revised manuscript, we have updated the abstract to include key metrics such as WER reduction percentages, medical term accuracy improvements, and explicit comparisons to baselines, along with references to the benchmark datasets used. This provides visible evidence for the improvements and SOTA positioning. revision: yes
Referee: [§3] §3 (KG construction): The framework relies on the medical KG supplying usable phonetic information (e.g., pronunciations, IPA mappings, or phonetic similarity metrics) alongside semantics. Standard resources such as UMLS contain semantic relations and synonyms but rarely encode phonetics; the manuscript must explicitly describe how phonetic features are added or derived, otherwise the correction step reduces to semantic lookup and cannot reliably address sound-alike medical-term errors.

Authors: We acknowledge that the original description of phonetic integration in §3 was insufficiently detailed. In the revised version, we have expanded §3 to explicitly explain the augmentation process: starting from UMLS for semantic relations, we derive phonetic features by mapping terms to IPA pronunciations via external resources such as the CMU Pronouncing Dictionary and computing similarity scores using phoneme-level edit distance. This ensures the KG supports both semantic and phonetic correction, distinguishing it from pure semantic lookup. revision: yes
Referee: [Experimental Section] Experimental evaluation: The manuscript must include quantitative results (e.g., WER or term-accuracy deltas), comparison tables against relevant baselines, statistical significance tests, and ablation studies on the phonetic component to substantiate the improvement and SOTA assertions.

Authors: We agree that the experimental section requires more comprehensive quantitative support to substantiate the claims. In the revised manuscript, we have added detailed tables reporting WER and term-accuracy deltas, comparisons against multiple baselines (including ASR-only, LLM-based correction, and KG-semantic-only variants), results of statistical significance tests (paired t-tests with p-values), and ablation studies that isolate the phonetic component's contribution. These additions directly address the need for rigorous evidence. revision: yes

Circularity Check

0 steps flagged

No circularity detected in derivation chain

full rationale

The paper introduces MedSpeak as an external-component framework that augments ASR transcripts via a medical knowledge graph (for semantic and phonetic relations) plus LLM reasoning. No quantity is defined in terms of itself, no fitted parameter is relabeled as a prediction, and no load-bearing premise reduces to a self-citation or prior ansatz by the same authors. The central claim rests on benchmark experiments rather than internal redefinition, rendering the derivation self-contained against external KG and LLM resources.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The framework rests on the assumption that an existing medical knowledge graph provides reliable phonetic and semantic links usable by LLMs for error correction; no free parameters or new entities are introduced in the abstract.

axioms (1)

domain assumption Medical knowledge graphs contain accurate semantic relationships and phonetic encodings for terminology that can correct ASR errors.
Invoked to justify the error correction mechanism.

pith-pipeline@v0.9.0 · 5436 in / 1059 out tokens · 21504 ms · 2026-05-16T09:22:17.128343+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

We construct our knowledge graph using the Unified Medical Language System (UMLS) ... use The CMU Pronouncing Dictionary to generate phonetically similar words ... Double Metaphone algorithm and Levenshtein Distance metric
IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

fine-tunes the parameters of a base LLM model under the full model supervision ... L(θ) = −∑ log P_θ(y_i | x, y_<i)

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

22 extracted references · 22 canonical work pages · 1 internal anchor

[1]

INTRODUCTION The increasing adoption of speech-based AI systems in medi- cal applications has led to the growing need for high-accuracy spoken medical question answering (SQA) systems. These systems rely on Automatic Speech Recognition (ASR) to transcribe spoken medical queries before feeding them into retrieval-augmented generation (RAG) models or large ...

work page internal anchor Pith review Pith/arXiv arXiv 2026
[2]

While fine-tuning ASR models on domain-specific datasets has been explored, these efforts are still limited by data scarcity and poor generalization [2, 7, 8]

RELATED WORKS Traditional ASR models, trained on general-domain corpora, often struggle to accurately recognize specialized medical ter- minology, leading to frequent misrecognitions of critical med- ical entities [1, 5, 6, 7]. While fine-tuning ASR models on domain-specific datasets has been explored, these efforts are still limited by data scarcity and ...

work page
[3]

MEDSPEAK Our MedSpeak framework combines static KG context injec- tion with a fine-tuned LLM, explicitly integrating both se- mantic relationship and phonetic similarity into LLM fine- tuning. This combined representation allows for robust ASR error correction and QA reasoning within a unified two-line supervision format, achieving great performance acros...

work page
[4]

We then evaluate our MedSpeak frame- work on spoken medical QA against various baselines

EXPERIMENT AND RESULTS In this section, we first describe the data preparation and ex- perimental settings. We then evaluate our MedSpeak frame- work on spoken medical QA against various baselines. 4.1. Data Preparation To systematically evaluate MedSpeak, we use a diverse med- ical SQA benchmark by synthesizing spoken data from three well-established mul...

work page 2048
[5]

CONCLUSION In this paper, we proposed MedSpeak, a framework integrates semantic and phonetic context from a medical knowledge graph with the reasoning ability of LLMs, enabling robust correction of transcription errors and reliable clinical reason- ing. Through extensive evaluation, MedSpeak demonstrated consistent improvements in both transcription accur...

work page
[6]

Zero-shot end-to-end spoken ques- tion answering in medical domain,

Yanis Labrak, Adel Moumen, Richard Dufour, and Mickael Rouvier, “Zero-shot end-to-end spoken ques- tion answering in medical domain,”arXiv preprint, 2025

work page 2025
[7]

Gec-rag: Improving generative error correction via retrieval-augmented generation for automatic speech recognition systems,

Amin Robatian, Mohammad Hajipour, Moham- mad Reza Peyghan, Fatemeh Rajabi, Sajjad Amini, Shahrokh Ghaemmaghami, and Iman Gholampour, “Gec-rag: Improving generative error correction via retrieval-augmented generation for automatic speech recognition systems,”arXiv preprint, 2025

work page 2025
[8]

Retrieval-augmented correction of named entity speech recognition errors,

Ernest Pusateri, Anmol Walia, Anirudh Kashi, Bor- tik Bandyopadhyay, Nadia Hyder, Sayantan Mahinder, Raviteja Anantha, Daben Liu, and Sashank Gondala, “Retrieval-augmented correction of named entity speech recognition errors,”IEEE, 2025

work page 2025
[9]

Retrieval- augmented end-to-end spoken dialog models,

Mingqiu Wang, Izhak Shafran, Hagen Soltau, Wei Han, Yuan Cao, Dian Yu, and Laurent El Shafey, “Retrieval- augmented end-to-end spoken dialog models,”ICASSP, 2025

work page 2025
[10]

Speech retrieval-augmented generation without auto- matic speech recognition,

Do June Min, Karel Mundnich, Andy Lapastora, Erfan Soltanmohammadi, Srikanth Ronanki, and Kyu Han, “Speech retrieval-augmented generation without auto- matic speech recognition,”ICASSP, 2025

work page 2025
[12]

Rasu: Retrieval-augmented speech understand- ing through generative modeling,

Hao Yang, Min Zhang, Minghan Wang, and Jiaxin Guo, “Rasu: Retrieval-augmented speech understand- ing through generative modeling,”IEEE, 2025

work page 2025
[13]

Seal: Speech embedding alignment learning for speech large language model with retrieval-augmented generation,

Chunyu Sun, Bingyu Liu, Zhichao Cui, Anbin Qi, Tian- Hao Zhang, Dinghao Zhou, and Lewei Lu, “Seal: Speech embedding alignment learning for speech large language model with retrieval-augmented generation,” arXiv preprint, 2025

work page 2025
[14]

Retrieval-augmented text-to-audio generation,

Yi Yuan, Haohe Liu, Xubo Liu, Qiushi Huang, Mark D. Plumbley, and Wenwu Wang, “Retrieval-augmented text-to-audio generation,”ICASSP, 2025

work page 2025
[15]

Contextualization of asr with llm using phonetic retrieval-based augmentation,

Zhihong Lei, Xingyu Na, Mingbin Xu, Ernest Pusateri, Christophe Van Gysel, Yuanyuan Zhang, Shiyi Han, and Zhen Huang, “Contextualization of asr with llm using phonetic retrieval-based augmentation,”arXiv preprint arXiv:2409.15353, 2024

work page arXiv 2024
[16]

Retrieval augmented correction of named entity speech recognition errors,

Ernest Pusateri, Anmol Walia, Anirudh Kashi, Bor- tik Bandyopadhyay, Nadia Hyder, Sayantan Mahinder, Raviteja Anantha, Daben Liu, and Sashank Gondala, “Retrieval augmented correction of named entity speech recognition errors,”arXiv preprint arXiv:2409.06062, 2024

work page arXiv 2024
[17]

Rasu: Retrieval augmented speech understand- ing through generative modeling,

Hao Yang, Min Zhang, Minghan Wang, and Jiaxin Guo, “Rasu: Retrieval augmented speech understand- ing through generative modeling,” inInterspeech, 2024, vol. 2024, pp. 3510–3514

work page 2024
[18]

Retrieval augmented generation in prompt- based text-to-speech synthesis with context-aware con- trastive language-audio pretraining,

Jinlong Xue, Yayue Deng, Yingming Gao, and Ya Li, “Retrieval augmented generation in prompt- based text-to-speech synthesis with context-aware con- trastive language-audio pretraining,”arXiv preprint arXiv:2406.03714, 2024

work page arXiv 2024
[19]

Contextual asr with retrieval augmented large language model,

Cihan Xiao, Zejiang Hou, Daniel Garcia-Romero, and Kyu J Han, “Contextual asr with retrieval augmented large language model,” inICASSP 2025-2025 IEEE In- ternational Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2025, pp. 1–5

work page 2025
[20]

Irag: Itera- tive retrieval-augmented generation for spoken language understanding,

Hao Yang, Min Zhang, and Daimeng Wei, “Irag: Itera- tive retrieval-augmented generation for spoken language understanding,”IEEE, 2025

work page 2025
[21]

Medrag: Enhancing retrieval-augmented generation with knowledge graph-elicited reasoning for healthcare copilot,

Xuejiao Zhao, Siyan Liu, Su-Yin Yang, and Chun- yan Miao, “Medrag: Enhancing retrieval-augmented generation with knowledge graph-elicited reasoning for healthcare copilot,”arXiv preprint, 2025

work page 2025
[22]

Unified medical lan- guage system (umls),

National Library of Medicine, “Unified medical lan- guage system (umls),” National Institutes of Health, 2024

work page 2024
[23]

Retrieval-augmented dialogue knowledge aggregation for expressive conversational speech synthesis,

Rui Liu, Zhenqi Jia, Feilong Bao, and Haizhou Li, “Retrieval-augmented dialogue knowledge aggregation for expressive conversational speech synthesis,”Infor- mation Fusion, 2025

work page 2025

[1] [1]

INTRODUCTION The increasing adoption of speech-based AI systems in medi- cal applications has led to the growing need for high-accuracy spoken medical question answering (SQA) systems. These systems rely on Automatic Speech Recognition (ASR) to transcribe spoken medical queries before feeding them into retrieval-augmented generation (RAG) models or large ...

work page internal anchor Pith review Pith/arXiv arXiv 2026

[2] [2]

While fine-tuning ASR models on domain-specific datasets has been explored, these efforts are still limited by data scarcity and poor generalization [2, 7, 8]

RELATED WORKS Traditional ASR models, trained on general-domain corpora, often struggle to accurately recognize specialized medical ter- minology, leading to frequent misrecognitions of critical med- ical entities [1, 5, 6, 7]. While fine-tuning ASR models on domain-specific datasets has been explored, these efforts are still limited by data scarcity and ...

work page

[3] [3]

MEDSPEAK Our MedSpeak framework combines static KG context injec- tion with a fine-tuned LLM, explicitly integrating both se- mantic relationship and phonetic similarity into LLM fine- tuning. This combined representation allows for robust ASR error correction and QA reasoning within a unified two-line supervision format, achieving great performance acros...

work page

[4] [4]

We then evaluate our MedSpeak frame- work on spoken medical QA against various baselines

EXPERIMENT AND RESULTS In this section, we first describe the data preparation and ex- perimental settings. We then evaluate our MedSpeak frame- work on spoken medical QA against various baselines. 4.1. Data Preparation To systematically evaluate MedSpeak, we use a diverse med- ical SQA benchmark by synthesizing spoken data from three well-established mul...

work page 2048

[5] [5]

CONCLUSION In this paper, we proposed MedSpeak, a framework integrates semantic and phonetic context from a medical knowledge graph with the reasoning ability of LLMs, enabling robust correction of transcription errors and reliable clinical reason- ing. Through extensive evaluation, MedSpeak demonstrated consistent improvements in both transcription accur...

work page

[6] [6]

Zero-shot end-to-end spoken ques- tion answering in medical domain,

Yanis Labrak, Adel Moumen, Richard Dufour, and Mickael Rouvier, “Zero-shot end-to-end spoken ques- tion answering in medical domain,”arXiv preprint, 2025

work page 2025

[7] [7]

Gec-rag: Improving generative error correction via retrieval-augmented generation for automatic speech recognition systems,

Amin Robatian, Mohammad Hajipour, Moham- mad Reza Peyghan, Fatemeh Rajabi, Sajjad Amini, Shahrokh Ghaemmaghami, and Iman Gholampour, “Gec-rag: Improving generative error correction via retrieval-augmented generation for automatic speech recognition systems,”arXiv preprint, 2025

work page 2025

[8] [8]

Retrieval-augmented correction of named entity speech recognition errors,

Ernest Pusateri, Anmol Walia, Anirudh Kashi, Bor- tik Bandyopadhyay, Nadia Hyder, Sayantan Mahinder, Raviteja Anantha, Daben Liu, and Sashank Gondala, “Retrieval-augmented correction of named entity speech recognition errors,”IEEE, 2025

work page 2025

[9] [9]

Retrieval- augmented end-to-end spoken dialog models,

Mingqiu Wang, Izhak Shafran, Hagen Soltau, Wei Han, Yuan Cao, Dian Yu, and Laurent El Shafey, “Retrieval- augmented end-to-end spoken dialog models,”ICASSP, 2025

work page 2025

[10] [10]

Speech retrieval-augmented generation without auto- matic speech recognition,

Do June Min, Karel Mundnich, Andy Lapastora, Erfan Soltanmohammadi, Srikanth Ronanki, and Kyu Han, “Speech retrieval-augmented generation without auto- matic speech recognition,”ICASSP, 2025

work page 2025

[11] [12]

Rasu: Retrieval-augmented speech understand- ing through generative modeling,

Hao Yang, Min Zhang, Minghan Wang, and Jiaxin Guo, “Rasu: Retrieval-augmented speech understand- ing through generative modeling,”IEEE, 2025

work page 2025

[12] [13]

Seal: Speech embedding alignment learning for speech large language model with retrieval-augmented generation,

Chunyu Sun, Bingyu Liu, Zhichao Cui, Anbin Qi, Tian- Hao Zhang, Dinghao Zhou, and Lewei Lu, “Seal: Speech embedding alignment learning for speech large language model with retrieval-augmented generation,” arXiv preprint, 2025

work page 2025

[13] [14]

Retrieval-augmented text-to-audio generation,

Yi Yuan, Haohe Liu, Xubo Liu, Qiushi Huang, Mark D. Plumbley, and Wenwu Wang, “Retrieval-augmented text-to-audio generation,”ICASSP, 2025

work page 2025

[14] [15]

Contextualization of asr with llm using phonetic retrieval-based augmentation,

Zhihong Lei, Xingyu Na, Mingbin Xu, Ernest Pusateri, Christophe Van Gysel, Yuanyuan Zhang, Shiyi Han, and Zhen Huang, “Contextualization of asr with llm using phonetic retrieval-based augmentation,”arXiv preprint arXiv:2409.15353, 2024

work page arXiv 2024

[15] [16]

Retrieval augmented correction of named entity speech recognition errors,

Ernest Pusateri, Anmol Walia, Anirudh Kashi, Bor- tik Bandyopadhyay, Nadia Hyder, Sayantan Mahinder, Raviteja Anantha, Daben Liu, and Sashank Gondala, “Retrieval augmented correction of named entity speech recognition errors,”arXiv preprint arXiv:2409.06062, 2024

work page arXiv 2024

[16] [17]

Rasu: Retrieval augmented speech understand- ing through generative modeling,

Hao Yang, Min Zhang, Minghan Wang, and Jiaxin Guo, “Rasu: Retrieval augmented speech understand- ing through generative modeling,” inInterspeech, 2024, vol. 2024, pp. 3510–3514

work page 2024

[17] [18]

Retrieval augmented generation in prompt- based text-to-speech synthesis with context-aware con- trastive language-audio pretraining,

Jinlong Xue, Yayue Deng, Yingming Gao, and Ya Li, “Retrieval augmented generation in prompt- based text-to-speech synthesis with context-aware con- trastive language-audio pretraining,”arXiv preprint arXiv:2406.03714, 2024

work page arXiv 2024

[18] [19]

Contextual asr with retrieval augmented large language model,

Cihan Xiao, Zejiang Hou, Daniel Garcia-Romero, and Kyu J Han, “Contextual asr with retrieval augmented large language model,” inICASSP 2025-2025 IEEE In- ternational Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2025, pp. 1–5

work page 2025

[19] [20]

Irag: Itera- tive retrieval-augmented generation for spoken language understanding,

Hao Yang, Min Zhang, and Daimeng Wei, “Irag: Itera- tive retrieval-augmented generation for spoken language understanding,”IEEE, 2025

work page 2025

[20] [21]

Medrag: Enhancing retrieval-augmented generation with knowledge graph-elicited reasoning for healthcare copilot,

Xuejiao Zhao, Siyan Liu, Su-Yin Yang, and Chun- yan Miao, “Medrag: Enhancing retrieval-augmented generation with knowledge graph-elicited reasoning for healthcare copilot,”arXiv preprint, 2025

work page 2025

[21] [22]

Unified medical lan- guage system (umls),

National Library of Medicine, “Unified medical lan- guage system (umls),” National Institutes of Health, 2024

work page 2024

[22] [23]

Retrieval-augmented dialogue knowledge aggregation for expressive conversational speech synthesis,

Rui Liu, Zhenqi Jia, Feilong Bao, and Haizhou Li, “Retrieval-augmented dialogue knowledge aggregation for expressive conversational speech synthesis,”Infor- mation Fusion, 2025

work page 2025