<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="bourrel-etal-2026-probe">
    <titleInfo>
        <title>A Probe Shift Is Not a Fairness Fix: The Limits of Representation Steering in Speech Models</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Nicolas</namePart>
        <namePart type="family">Bourrel</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Abderrahmane</namePart>
        <namePart type="family">Issam</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Gerasimos</namePart>
        <namePart type="family">Spanakis</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>Automatic speech recognition (ASR) systems exhibit unequal error rates across speaker groups, motivating interventions on their internal representations. We ask whether speakerlinked attributes that are linearly readable from pretrained ASR encoders yield useful directions for reducing group word-error-rate (WER) gaps. Across Whisper-medium, HuBERT-large, and Wav2Vec2-large on Common Voice and the Speech Accent Archive, we probe every encoder layer for metadata-derived sex/gender, age, and native/accent labels; construct centroid and probe-derived directions; inject them at selected layers; and compare downstream probe trajectories with matched WER changes. Sex labels are highly decodable (best macroF1 0.924–0.941), native/accent labels are also above chance (0.544–0.696), and age is weaker (0.354–0.397). Of 22 post-selected reruns, nine have 95% paired-bootstrap intervals entirely below zero, yet every absolute source-group WER reduction is below 0.7 percentage points. Conversely, a local target-class probe rate can rise from 8.09% to 99.87% while WER worsens. Linear readability is therefore neither evidence of causal use nor a reliable mitigation method. Our results motivate evaluating speech-bias interventions jointly at representation, propagation, and task levels.</abstract>
    <identifier type="citekey">bourrel-etal-2026-probe</identifier>
    <location>
        <url>https://aclanthology.org/2026.impactspeech-1.1/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>1</start>
            <end>9</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T A Probe Shift Is Not a Fairness Fix: The Limits of Representation Steering in Speech Models
%A Bourrel, Nicolas
%A Issam, Abderrahmane
%A Spanakis, Gerasimos
%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 bourrel-etal-2026-probe
%X Automatic speech recognition (ASR) systems exhibit unequal error rates across speaker groups, motivating interventions on their internal representations. We ask whether speakerlinked attributes that are linearly readable from pretrained ASR encoders yield useful directions for reducing group word-error-rate (WER) gaps. Across Whisper-medium, HuBERT-large, and Wav2Vec2-large on Common Voice and the Speech Accent Archive, we probe every encoder layer for metadata-derived sex/gender, age, and native/accent labels; construct centroid and probe-derived directions; inject them at selected layers; and compare downstream probe trajectories with matched WER changes. Sex labels are highly decodable (best macroF1 0.924–0.941), native/accent labels are also above chance (0.544–0.696), and age is weaker (0.354–0.397). Of 22 post-selected reruns, nine have 95% paired-bootstrap intervals entirely below zero, yet every absolute source-group WER reduction is below 0.7 percentage points. Conversely, a local target-class probe rate can rise from 8.09% to 99.87% while WER worsens. Linear readability is therefore neither evidence of causal use nor a reliable mitigation method. Our results motivate evaluating speech-bias interventions jointly at representation, propagation, and task levels.
%U https://aclanthology.org/2026.impactspeech-1.1/
%P 1-9

Markdown (Informal)

[A Probe Shift Is Not a Fairness Fix: The Limits of Representation Steering in Speech Models](https://aclanthology.org/2026.impactspeech-1.1/) (Bourrel et al., IMPACT-SPEECH 2026)

• A Probe Shift Is Not a Fairness Fix: The Limits of Representation Steering in Speech Models (Bourrel et al., IMPACT-SPEECH 2026)

ACL

• Nicolas Bourrel, Abderrahmane Issam, and Gerasimos Spanakis. 2026. A Probe Shift Is Not a Fairness Fix: The Limits of Representation Steering in Speech Models. In Proceedings of the IMPACT-SPEECH Workshop: Identifying, Measuring, Preventing, and Assessing Consequences of Bias in Speech LLMs, pages 1–9, Budapest, Hungary. Association for Computational Linguistics.