@inproceedings{pan-etal-2026-whisperx,
title = "{W}hisper{X}-{LLM}-{T}erm{F}ixer: Enhancing Whisper with {LLM}-Guided Post-Decoding for Domain-Specific Terminology",
author = "Pan, Ronghao and
Garc{\'i}a-D{\'i}az, Jos{\'e} Antonio and
Valencia-Garc{\'i}a, Rafael",
editor = "Zhang, Yue and
Xu, Ruifeng and
Purwarianti, Ayu and
Rong, Tong and
Suzuki, Jun",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {IJCNLP}-{AACL} 2026",
month = nov,
year = "2026",
address = "Hengqin, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-aacl.153/",
pages = "3019--3042",
ISBN = "979-8-89176-444-6",
abstract = "Large-scale ASR models such as Whisper perform well in general domains but struggle with specialized terminology, particularly in sensitive domains like medicine. We propose WhisperX-LLM-TermFixer, a lightweight post-decoding framework that improves term recognition without retraining or modifying the acoustic model. The method combines N-best decoding, phonetic n-gram correction, and LLM-based contextual filtering. We evaluate the method on a controlled synthetic benchmark containing 1,005 utterances per language in English, Spanish, Catalan, and Basque, with one target medical term per utterance and full lexicon coverage. Across six Whisper sizes, the selected configurations improve term accuracy by 16.12-35.32 percentage points and reduce WER by 0.88-4.01 points relative to WhisperX. LLM filtering reduces WER relative to the phonetic stage in all evaluated configurations, although it reduces term accuracy in three Basque configurations. Perplexity results provide additional evidence of improved linguistic plausibility. The selected configurations require approximately 1.26x-2.38x the baseline inference time. Our code and data are publicly available at https://github.com/NLP-UMUTeam/WhisperX-LLM-TermFixer."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="pan-etal-2026-whisperx">
<titleInfo>
<title>WhisperX-LLM-TermFixer: Enhancing Whisper with LLM-Guided Post-Decoding for Domain-Specific Terminology</title>
</titleInfo>
<name type="personal">
<namePart type="given">Ronghao</namePart>
<namePart type="family">Pan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">José</namePart>
<namePart type="given">Antonio</namePart>
<namePart type="family">García-Díaz</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Rafael</namePart>
<namePart type="family">Valencia-García</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-11</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yue</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ruifeng</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ayu</namePart>
<namePart type="family">Purwarianti</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tong</namePart>
<namePart type="family">Rong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jun</namePart>
<namePart type="family">Suzuki</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Hengqin, China</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-444-6</identifier>
</relatedItem>
<abstract>Large-scale ASR models such as Whisper perform well in general domains but struggle with specialized terminology, particularly in sensitive domains like medicine. We propose WhisperX-LLM-TermFixer, a lightweight post-decoding framework that improves term recognition without retraining or modifying the acoustic model. The method combines N-best decoding, phonetic n-gram correction, and LLM-based contextual filtering. We evaluate the method on a controlled synthetic benchmark containing 1,005 utterances per language in English, Spanish, Catalan, and Basque, with one target medical term per utterance and full lexicon coverage. Across six Whisper sizes, the selected configurations improve term accuracy by 16.12-35.32 percentage points and reduce WER by 0.88-4.01 points relative to WhisperX. LLM filtering reduces WER relative to the phonetic stage in all evaluated configurations, although it reduces term accuracy in three Basque configurations. Perplexity results provide additional evidence of improved linguistic plausibility. The selected configurations require approximately 1.26x-2.38x the baseline inference time. Our code and data are publicly available at https://github.com/NLP-UMUTeam/WhisperX-LLM-TermFixer.</abstract>
<identifier type="citekey">pan-etal-2026-whisperx</identifier>
<location>
<url>https://aclanthology.org/2026.findings-aacl.153/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>3019</start>
<end>3042</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T WhisperX-LLM-TermFixer: Enhancing Whisper with LLM-Guided Post-Decoding for Domain-Specific Terminology
%A Pan, Ronghao
%A García-Díaz, José Antonio
%A Valencia-García, Rafael
%Y Zhang, Yue
%Y Xu, Ruifeng
%Y Purwarianti, Ayu
%Y Rong, Tong
%Y Suzuki, Jun
%S Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026
%D 2026
%8 November
%I Association for Computational Linguistics
%C Hengqin, China
%@ 979-8-89176-444-6
%F pan-etal-2026-whisperx
%X Large-scale ASR models such as Whisper perform well in general domains but struggle with specialized terminology, particularly in sensitive domains like medicine. We propose WhisperX-LLM-TermFixer, a lightweight post-decoding framework that improves term recognition without retraining or modifying the acoustic model. The method combines N-best decoding, phonetic n-gram correction, and LLM-based contextual filtering. We evaluate the method on a controlled synthetic benchmark containing 1,005 utterances per language in English, Spanish, Catalan, and Basque, with one target medical term per utterance and full lexicon coverage. Across six Whisper sizes, the selected configurations improve term accuracy by 16.12-35.32 percentage points and reduce WER by 0.88-4.01 points relative to WhisperX. LLM filtering reduces WER relative to the phonetic stage in all evaluated configurations, although it reduces term accuracy in three Basque configurations. Perplexity results provide additional evidence of improved linguistic plausibility. The selected configurations require approximately 1.26x-2.38x the baseline inference time. Our code and data are publicly available at https://github.com/NLP-UMUTeam/WhisperX-LLM-TermFixer.
%U https://aclanthology.org/2026.findings-aacl.153/
%P 3019-3042
Markdown (Informal)
[WhisperX-LLM-TermFixer: Enhancing Whisper with LLM-Guided Post-Decoding for Domain-Specific Terminology](https://aclanthology.org/2026.findings-aacl.153/) (Pan et al., Findings 2026)
ACL