@inproceedings{park-kim-2026-acoustic,
title = "Not All Acoustic Frames Contribute Equally: Frame-Selective Acoustic Conditioning for Dysarthric {ASR} Correction",
author = "Park, Gyutae and
Kim, Ji-Hwan",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.409/",
pages = "8638--8657",
ISBN = "979-8-89176-462-0",
abstract = "Not all acoustic frames are equally useful for correcting dysarthric ASR errors. In dysarthric speech, acoustic deviations are unevenly distributed across frames, but LLM-based generative error correction systems often inject acoustic features uniformly, under-emphasizing deviant frames and over-correcting reliable ones. We propose a frame-selective acoustic conditioning method driven by a Frame Deviation Score (FDS) and VGate. FDS is a control-referenced per-frame deviation score computed over Whisper encoder frames, and VGate uses standardized per-frame FDS to reweight acoustic value vectors. Across ten personalized TORGO settings, we compare VGate with Plain, an acoustic-adapter baseline without the VGate gating module, and show that VGate is conditional rather than universally beneficial. Three findings support conditional use. (1) Plain residual headroom (the gap between Plain and Oracle@5 WER) and utterance-mean FDS-DIV characterize high-benefit cases, with gains concentrated in ref-preserved and overgeneration cases and the largest improvements driven by insertion control. (2) A development-calibrated FDS-DIV router lowers pooled WER from 19.55{\%} to 18.29{\%} while routing only 20.72{\%} of utterances. (3) Cross-corpus COPAS results and blinded listening further support this conditional-use strategy. Our code is publicly available at https://github.com/doraemon500/FSAC."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="park-kim-2026-acoustic">
<titleInfo>
<title>Not All Acoustic Frames Contribute Equally: Frame-Selective Acoustic Conditioning for Dysarthric ASR Correction</title>
</titleInfo>
<name type="personal">
<namePart type="given">Gyutae</namePart>
<namePart type="family">Park</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ji-Hwan</namePart>
<namePart type="family">Kim</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EMNLP 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>Not all acoustic frames are equally useful for correcting dysarthric ASR errors. In dysarthric speech, acoustic deviations are unevenly distributed across frames, but LLM-based generative error correction systems often inject acoustic features uniformly, under-emphasizing deviant frames and over-correcting reliable ones. We propose a frame-selective acoustic conditioning method driven by a Frame Deviation Score (FDS) and VGate. FDS is a control-referenced per-frame deviation score computed over Whisper encoder frames, and VGate uses standardized per-frame FDS to reweight acoustic value vectors. Across ten personalized TORGO settings, we compare VGate with Plain, an acoustic-adapter baseline without the VGate gating module, and show that VGate is conditional rather than universally beneficial. Three findings support conditional use. (1) Plain residual headroom (the gap between Plain and Oracle@5 WER) and utterance-mean FDS-DIV characterize high-benefit cases, with gains concentrated in ref-preserved and overgeneration cases and the largest improvements driven by insertion control. (2) A development-calibrated FDS-DIV router lowers pooled WER from 19.55% to 18.29% while routing only 20.72% of utterances. (3) Cross-corpus COPAS results and blinded listening further support this conditional-use strategy. Our code is publicly available at https://github.com/doraemon500/FSAC.</abstract>
<identifier type="citekey">park-kim-2026-acoustic</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.409/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>8638</start>
<end>8657</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Not All Acoustic Frames Contribute Equally: Frame-Selective Acoustic Conditioning for Dysarthric ASR Correction
%A Park, Gyutae
%A Kim, Ji-Hwan
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F park-kim-2026-acoustic
%X Not all acoustic frames are equally useful for correcting dysarthric ASR errors. In dysarthric speech, acoustic deviations are unevenly distributed across frames, but LLM-based generative error correction systems often inject acoustic features uniformly, under-emphasizing deviant frames and over-correcting reliable ones. We propose a frame-selective acoustic conditioning method driven by a Frame Deviation Score (FDS) and VGate. FDS is a control-referenced per-frame deviation score computed over Whisper encoder frames, and VGate uses standardized per-frame FDS to reweight acoustic value vectors. Across ten personalized TORGO settings, we compare VGate with Plain, an acoustic-adapter baseline without the VGate gating module, and show that VGate is conditional rather than universally beneficial. Three findings support conditional use. (1) Plain residual headroom (the gap between Plain and Oracle@5 WER) and utterance-mean FDS-DIV characterize high-benefit cases, with gains concentrated in ref-preserved and overgeneration cases and the largest improvements driven by insertion control. (2) A development-calibrated FDS-DIV router lowers pooled WER from 19.55% to 18.29% while routing only 20.72% of utterances. (3) Cross-corpus COPAS results and blinded listening further support this conditional-use strategy. Our code is publicly available at https://github.com/doraemon500/FSAC.
%U https://aclanthology.org/2026.findings-emnlp.409/
%P 8638-8657
Markdown (Informal)
[Not All Acoustic Frames Contribute Equally: Frame-Selective Acoustic Conditioning for Dysarthric ASR Correction](https://aclanthology.org/2026.findings-emnlp.409/) (Park & Kim, Findings 2026)
ACL