@inproceedings{bajwa-2026-lost,
title = "Lost in Compression: Calibration Bias and the Accent-{WER} Gap in Speech {LLM}s",
author = "Bajwa, Angad Ripudaman Singh",
editor = "Shekhar, Ravi and
Swain, Monorama and
Mishra, Jagabandhu and
Dhar, Sandipan and
Mouratidis, Haralambos and
Purver, Matthew",
booktitle = "Proceedings of the {IMPACT}-{SPEECH} Workshop: Identifying, Measuring, Preventing, and Assessing Consequences of Bias in Speech {LLM}s",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.impactspeech-1.6/",
pages = "61--66",
ISBN = "979-8-89176-479-8",
abstract = "GPTQ and similar calibration-based quantization methods tune their rounding to a small calibration sample, in practice usually convenient, native-accented read speech. Edge devices disproportionately run quantized models, so the speakers most likely to depend on the compressed model are often the ones calibration represented least. Compression is known to amplify bias in text LLMs and across languages in multilingual speech models, but whether it does so within a single language, between native and non-native accents, in a modern speech LLM, remains untested. We quantize the LLM backbone of Qwen2-Audio-7B-Instruct with GPTQ, hold the recipe fixed and vary only which EdAcc speakers fill the calibration set (native-only, balanced, L2-only), measuring the native-vs-L2 word error rate (WER) gap at matched overall WER. We expected nativeonly calibration to widen the gap and balanced calibration to close it. Instead, homogeneous calibration, native-only or L2-only, narrows the gap while balanced calibration does almost nothing. It recurs when we quantize the audio encoder, weakens at 3-bit and collapses at 2-bit."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="bajwa-2026-lost">
<titleInfo>
<title>Lost in Compression: Calibration Bias and the Accent-WER Gap in Speech LLMs</title>
</titleInfo>
<name type="personal">
<namePart type="given">Angad</namePart>
<namePart type="given">Ripudaman</namePart>
<namePart type="given">Singh</namePart>
<namePart type="family">Bajwa</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>Proceedings of the IMPACT-SPEECH Workshop: Identifying, Measuring, Preventing, and Assessing Consequences of Bias in Speech LLMs</title>
</titleInfo>
<name type="personal">
<namePart type="given">Ravi</namePart>
<namePart type="family">Shekhar</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Monorama</namePart>
<namePart type="family">Swain</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jagabandhu</namePart>
<namePart type="family">Mishra</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sandipan</namePart>
<namePart type="family">Dhar</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Haralambos</namePart>
<namePart type="family">Mouratidis</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Matthew</namePart>
<namePart type="family">Purver</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">979-8-89176-479-8</identifier>
</relatedItem>
<abstract>GPTQ and similar calibration-based quantization methods tune their rounding to a small calibration sample, in practice usually convenient, native-accented read speech. Edge devices disproportionately run quantized models, so the speakers most likely to depend on the compressed model are often the ones calibration represented least. Compression is known to amplify bias in text LLMs and across languages in multilingual speech models, but whether it does so within a single language, between native and non-native accents, in a modern speech LLM, remains untested. We quantize the LLM backbone of Qwen2-Audio-7B-Instruct with GPTQ, hold the recipe fixed and vary only which EdAcc speakers fill the calibration set (native-only, balanced, L2-only), measuring the native-vs-L2 word error rate (WER) gap at matched overall WER. We expected nativeonly calibration to widen the gap and balanced calibration to close it. Instead, homogeneous calibration, native-only or L2-only, narrows the gap while balanced calibration does almost nothing. It recurs when we quantize the audio encoder, weakens at 3-bit and collapses at 2-bit.</abstract>
<identifier type="citekey">bajwa-2026-lost</identifier>
<location>
<url>https://aclanthology.org/2026.impactspeech-1.6/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>61</start>
<end>66</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Lost in Compression: Calibration Bias and the Accent-WER Gap in Speech LLMs
%A Bajwa, Angad Ripudaman Singh
%Y Shekhar, Ravi
%Y Swain, Monorama
%Y Mishra, Jagabandhu
%Y Dhar, Sandipan
%Y Mouratidis, Haralambos
%Y Purver, Matthew
%S Proceedings of the IMPACT-SPEECH Workshop: Identifying, Measuring, Preventing, and Assessing Consequences of Bias in Speech LLMs
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-479-8
%F bajwa-2026-lost
%X GPTQ and similar calibration-based quantization methods tune their rounding to a small calibration sample, in practice usually convenient, native-accented read speech. Edge devices disproportionately run quantized models, so the speakers most likely to depend on the compressed model are often the ones calibration represented least. Compression is known to amplify bias in text LLMs and across languages in multilingual speech models, but whether it does so within a single language, between native and non-native accents, in a modern speech LLM, remains untested. We quantize the LLM backbone of Qwen2-Audio-7B-Instruct with GPTQ, hold the recipe fixed and vary only which EdAcc speakers fill the calibration set (native-only, balanced, L2-only), measuring the native-vs-L2 word error rate (WER) gap at matched overall WER. We expected nativeonly calibration to widen the gap and balanced calibration to close it. Instead, homogeneous calibration, native-only or L2-only, narrows the gap while balanced calibration does almost nothing. It recurs when we quantize the audio encoder, weakens at 3-bit and collapses at 2-bit.
%U https://aclanthology.org/2026.impactspeech-1.6/
%P 61-66
Markdown (Informal)
[Lost in Compression: Calibration Bias and the Accent-WER Gap in Speech LLMs](https://aclanthology.org/2026.impactspeech-1.6/) (Bajwa, IMPACT-SPEECH 2026)
ACL