pith. machine review for the scientific record. sign in

arxiv: 2604.13067 · v1 · submitted 2026-03-19 · 💻 cs.HC · cs.CL

Recognition: no theorem link

From Seeing it to Experiencing it: Interactive Evaluation of Intersectional Voice Bias in Human-AI Speech Interaction

Shree Harsha Bokkahalli Satish , Maria Teleki , Christoph Minixhofer , Ondrej Klejch , Peter Bell , \'Eva Sz\'ekely
Authors on Pith no claims yet

Pith reviewed 2026-05-15 07:56 UTC · model grok-4.3

classification 💻 cs.HC cs.CL
keywords intersectional biasvoice conversionSpeechLLMsaccent biasgender biasinteractive evaluationquality-of-service disparitieshuman-AI speech interaction
0
0 comments X

The pith

Voice conversion in interactive studies increases user trust in AI speech responses and highlights intersectional accent and gender biases.

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

The paper proposes a two-part evaluation to better understand bias in SpeechLLMs: automated analysis for quality-of-service issues and interactive user studies using voice conversion. It aims to show how users experience bias when they can hear the same content in different voices. The results indicate that experiencing responses through varied vocal identities boosts trust and encourages users to consider different perspectives. Automated checks across models reveal differences in how well responses align with prompts and their length, depending on the combination of accent and perceived gender.

Core claim

By combining judge-free metrics with an interactive setup where voice conversion lets participants hear identical AI outputs voiced as different accents and genders, the work finds that this experience raises trust and acceptability for neutral responses while promoting perspective-taking, and that automated metrics detect accent-by-gender differences in response alignment and verbosity.

What carries the argument

Voice conversion to let users experience the same content through different vocal identities, paired with automated prompt-response metrics for alignment and verbosity to detect quality-of-service disparities.

Load-bearing premise

The key assumption is that changing a voice to sound like a different accent or gender does not introduce other changes that affect how people judge the AI's response, and that the automatic measurements fully show the service quality differences that users experience.

What would settle it

A test where participants rate the naturalness of the voice-converted audio samples independently of the content to check for confounding artifacts, or where the automated alignment and verbosity scores are compared directly to human judgments of response quality.

Figures

Figures reproduced from arXiv: 2604.13067 by Christoph Minixhofer, \'Eva Sz\'ekely, Maria Teleki, Ondrej Klejch, Peter Bell, Shree Harsha Bokkahalli Satish.

Figure 1
Figure 1. Figure 1: Study design – Observational Study (OS) and Interactive Study (IS) – illustrating three positions from which participants [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The same inputs are treated differently by the SpeechLLM depending on [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of Harm, Acceptability, and Trust ratings across conditions for the [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
read the original abstract

SpeechLLMs process spoken language directly from audio, but accent and vocal identity cues can lead to biased behaviour. Current bias evaluations often miss how such bias manifests in end-to-end speech interactions and how users experience it. We distinguish quality-of-service disparities (e.g., off-topic or low-effort responses) from content-level bias in coherent outputs, and examine intersectional effects of accent and perceived gender. In this work, we explore a two-part evaluation approach: (1) a controlled test cohort spanning six accents and two gender presentations, analysed with judge-free prompt-response metrics, and (2) an interactive study design using voice conversion to let users experience identical content through different vocal identities. Across two studies (Interactive, N=24; Observational, N=19), we find that voice conversion increases trust and acceptability for benign responses and encourages perspective-taking, while automated analysis in search of quality-of-service disparities, reveals {accent x gender} disparities in alignment and verbosity across SpeechLLMs. These results highlight voice conversion for probing and experiencing intersectional voice bias while our evaluation suite provides richer bias evaluations for spoken conversational AI.

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.

Referee Report

3 major / 2 minor

Summary. The paper claims that SpeechLLMs exhibit intersectional voice bias (accent x perceived gender) in end-to-end interactions, shown via (1) automated judge-free metrics revealing disparities in alignment and verbosity across a six-accent, two-gender cohort and (2) an interactive study (N=24) where voice conversion of identical benign responses increases user trust, acceptability, and perspective-taking, with a supporting observational study (N=19).

Significance. If the causal claims hold after addressing confounds, the work offers a novel interactive paradigm for users to experience rather than merely observe voice bias, plus a practical judge-free evaluation suite that separates quality-of-service disparities from content bias; this could meaningfully advance bias auditing for conversational speech AI beyond static text or single-accent tests.

major comments (3)
  1. [Interactive Study / Methods] Interactive study (N=24): the central claim that voice conversion increases trust/acceptability and encourages perspective-taking assumes converted voices isolate accent/gender identity without audible artifacts (unnatural prosody, timbre shifts, or intelligibility loss) that independently drive ratings; the manuscript appears to lack a control condition reporting naturalness/clarity ratings or artifact detection on the stimuli, directly threatening causal interpretation of the trust differences.
  2. [Observational Study / Automated Metrics] Automated analysis section: the reported {accent x gender} disparities in alignment and verbosity rest on judge-free prompt-response metrics whose exact definitions, aggregation rules, and statistical tests (including multiple-comparison correction and effect sizes) are not detailed in the abstract or visible methods summary; without these, it is impossible to verify whether the observed differences support the quality-of-service bias claims.
  3. [Results] Results and discussion: both studies use modest samples (N=24 and N=19) without reported power analysis or pre-registration; this weakens the generalizability of the intersectional disparity and perspective-taking findings, especially given the free parameters in accent/gender selection.
minor comments (2)
  1. [Abstract] The notation {accent x gender} is non-standard and should be defined explicitly on first use (e.g., 'accent-by-gender interaction').
  2. [Figures] Figure captions and axis labels for any automated metric plots should include the precise metric formulas and sample sizes per cell to improve readability.

Simulated Author's Rebuttal

3 responses · 0 unresolved

Thank you for the detailed and constructive review. We appreciate the referee's focus on strengthening causal claims in the interactive study, clarifying the automated metrics, and addressing sample size limitations. We address each major comment below and will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: [Interactive Study / Methods] Interactive study (N=24): the central claim that voice conversion increases trust/acceptability and encourages perspective-taking assumes converted voices isolate accent/gender identity without audible artifacts (unnatural prosody, timbre shifts, or intelligibility loss) that independently drive ratings; the manuscript appears to lack a control condition reporting naturalness/clarity ratings or artifact detection on the stimuli, directly threatening causal interpretation of the trust differences.

    Authors: We acknowledge this concern as valid for strengthening causal interpretation. Our voice conversion used a high-quality model selected to minimize artifacts while preserving content, but the manuscript does not report explicit naturalness or clarity ratings. In the revision, we will add a supplementary control analysis with participant ratings of naturalness, clarity, and artifact detection on the stimuli to confirm that trust differences are attributable to vocal identity rather than audio quality. revision: yes

  2. Referee: [Observational Study / Automated Metrics] Automated analysis section: the reported {accent x gender} disparities in alignment and verbosity rest on judge-free prompt-response metrics whose exact definitions, aggregation rules, and statistical tests (including multiple-comparison correction and effect sizes) are not detailed in the abstract or visible methods summary; without these, it is impossible to verify whether the observed differences support the quality-of-service bias claims.

    Authors: The full methods section defines alignment as cosine similarity between prompt and response embeddings (using a fixed sentence transformer) and verbosity as response token count, with per-condition means and two-way ANOVA followed by post-hoc tests. To improve transparency as requested, the revision will expand the methods with exact formulas, aggregation procedures, confirmation of Bonferroni correction, and reporting of effect sizes (Cohen's d and partial eta-squared) alongside the results. revision: yes

  3. Referee: [Results] Results and discussion: both studies use modest samples (N=24 and N=19) without reported power analysis or pre-registration; this weakens the generalizability of the intersectional disparity and perspective-taking findings, especially given the free parameters in accent/gender selection.

    Authors: We agree that the exploratory sample sizes limit strong generalizability claims and that power analysis and pre-registration would be preferable. Pre-registration was not used for this initial work. In the revision, we will add a post-hoc power analysis based on observed effects, expand the limitations section to discuss generalizability constraints and accent/gender selection choices, and recommend pre-registration for follow-up studies with larger samples. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical results from user studies and automated metrics

full rationale

The paper reports findings from two empirical studies (interactive N=24 and observational N=19) using voice conversion stimuli and judge-free prompt-response metrics to measure trust, acceptability, alignment, and verbosity. No equations, fitted parameters, self-referential definitions, or load-bearing self-citations appear in the derivation chain; claims derive directly from collected participant ratings and automated counts rather than reducing to prior inputs by construction. The evaluation is self-contained against external benchmarks of user perception and model output analysis.

Axiom & Free-Parameter Ledger

1 free parameters · 0 axioms · 0 invented entities

Evaluation depends on standard HCI assumptions about user perception and metric validity rather than new postulates.

free parameters (1)
  • selection of six accents and two gender presentations
    Chosen to span the test cohort; specific selection criteria not detailed in abstract.

pith-pipeline@v0.9.0 · 5529 in / 1053 out tokens · 55446 ms · 2026-05-15T07:56:19.833409+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

32 extracted references · 32 canonical work pages

  1. [1]

    2025.Amazon Alexa Echo

    Amazon.com. 2025.Amazon Alexa Echo. https://www.amazon.com/b?ie=UTF8& node=210779651011

    work page 2025

  2. [2]

    John Baugh. 2003. Linguistic profiling.Black linguistics: Language, society, and politics in Africa and the Americas(2003), 155–168

    work page 2003

  3. [3]

    Shree Harsha Bokkahalli Satish, Gustav Eje Henter, and Éva Székely. 2025. Do Bias Benchmarks Generalise? Evidence from Voice-based Evaluation of Gender Bias in SpeechLLMs.arXiv preprint arXiv:2510.01254(2025)

    work page arXiv 2025

  4. [4]

    Vineeta Chand. 2009. Social evaluations of accented Englishes: An Indian per- spective.Studies in the Linguistic Sciences: Illinois Working Papers(2009), 90–113

    work page 2009

  5. [5]

    Anna Seo Gyeong Choi and Hoon Choi. 2025. Fairness of Automatic Speech Recognition: Looking Through a Philosophical Lens. arXiv:2508.07143 [cs.CL] https://arxiv.org/abs/2508.07143

    work page arXiv 2025

  6. [6]

    Wenqian Cui, Dianzhi Yu, Xiaoqi Jiao, Ziqiao Meng, Guangyan Zhang, Qichao Wang, Yiwen Guo, and Irwin King. 2025. Recent advances in speech language models: a survey. doi:10.48550/arXiv.2410.03751 arXiv:2410.03751 [cs]

    work page doi:10.48550/arxiv.2410.03751 2025

  7. [7]

    Jay L Cunningham, Adinawa Adjagbodjou, Jeffrey Basoah, Jainaba Jawara, Kowe Kadoma, and Aaleyah Lewis. 2025. Toward Responsible ASR for African American English Speakers: A Scoping Review of Bias and Equity in Speech Technology. arXiv preprint arXiv:2508.18288(2025)

    work page arXiv 2025

  8. [8]

    Siyuan Feng, Bence Mark Halpern, Olya Kudina, and Odette Scharenborg. 2024. Towards inclusive automatic speech recognition.Computer Speech & Language 84 (2024), 101567

    work page 2024

  9. [9]

    2025.Gemini

    Google. 2025.Gemini. https://gemini.google/us/assistant/?hl=en

    work page 2025

  10. [10]

    Google. 2025. Manage Your Smart Home With Google Home. https://home. google.com/welcome/

    work page 2025

  11. [11]

    Camille Harris, Chijioke Mgbahurike, Neha Kumar, and Diyi Yang. 2024. Modeling Gender and Dialect Bias in Automatic Speech Recognition. InFindings of the Association for Computational Linguistics: EMNLP 2024, Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen (Eds.). Association for Computational Linguistics, Miami, Florida, USA, 15166–15184. doi:10.18653/...

    work page doi:10.18653/v1/2024.findings-emnlp.890 2024

  12. [12]

    Ivona Hideg and Laura J Kray. 2024. Speaking with a non-native accent: Impli- cations for work and personal life.Annual Review of Organizational Psychology and Organizational Behavior11 (2024), 109–134

    work page 2024

  13. [13]

    Valentin Hofmann, Pratyusha R Kalluri, Dan Jurafsky, and Sharese King. 2024. Dialect prejudice predicts AI decisions about people’s character, employability, and criminality.Nature633, 8029 (2024), 355–362

    work page 2024

  14. [14]

    Ziyue Jiang, Yi Ren, Ruiqi Li, Shengpeng Ji, Boyang Zhang, Zhenhui Ye, Chen Zhang, Bai Jionghao, Xiaoda Yang, Jialong Zuo, et al. 2025. MegaTTS 3: Sparse alignment enhanced latent diffusion transformer for zero-shot speech synthesis. arXiv preprint arXiv:2502.18924(2025)

    work page arXiv 2025

  15. [15]

    Allison Koenecke, Andrew Nam, Emily Lake, Joe Nudell, Minnie Quartey, Zion Mengesha, Connor Toups, John R Rickford, Dan Jurafsky, and Sharad Goel. 2020. Racial disparities in automated speech recognition.Proceedings of the national academy of sciences117, 14 (2020), 7684–7689

    work page 2020

  16. [16]

    Harm Lameris, Shree Harsha Bokkahalli Satish, Joakim Gustafson, and Éva Székely. 2025. Lost in Phonation: Voice Quality Variation as an Evaluation Dimension for Speech Foundation Models.arXiv preprint arXiv:2510.25577(2025)

    work page arXiv 2025

  17. [17]

    Shiri Lev-Ari and Boaz Keysar. 2010. Understanding of others’ mental states affects language processing in comprehension.Psychological Science21, 7 (2010), 1095–1102

    work page 2010

  18. [18]

    Songting Liu. 2024. Zero-shot voice conversion with diffusion transformers. arXiv preprint arXiv:2411.09943(2024)

    work page arXiv 2024

  19. [19]

    Jackson G Lu, Richard E Nisbett, and Michael W Morris. 2020. The bamboo ceiling: Exploring social psychological mechanisms for the underrepresentation of Asian Americans in leadership positions.Proceedings of the National Academy of Sciences117, 9 (2020), 4590–4596

    work page 2020

  20. [20]

    I don’t think these devices are very culturally sensi- tive

    Zion Mengesha, Courtney Heldreth, Michal Lahav, Juliana Sublewski, and El- yse Tuennerman. 2021. “I don’t think these devices are very culturally sensi- tive. ”—Impact of automated speech recognition errors on African Americans. Frontiers in Artificial Intelligence4 (2021), 725911

    work page 2021

  21. [21]

    Meta. 2025. Meta AI Glasses. https://www.meta.com/ai-glasses/

    work page 2025

  22. [22]

    Cade Metz. 2020. There Is a Racial Divide in Speech-Recognition Systems, Re- searchers Say.The New York Times(23 March 2020). https://www.nytimes.com/ 2020/03/23/technology/speech-recognition-bias-apple-amazon-google.html

    work page 2020

  23. [23]

    2025.Voice conversations in ChatGPT

    OpenAI. 2025.Voice conversations in ChatGPT. https://chatgpt.com/features/ voice/

    work page 2025

  24. [24]

    Thomas Purnell, William Idsardi, and John Baugh. 1999. Perceptual and phonetic experiments on American English dialect identification.Journal of Language and Social Psychology18, 1 (1999), 10–30

    work page 1999

  25. [25]

    Ramon Sanabria, Nikolay Bogoychev, Nina Markl, Andrea Carmantini, Ondrej Klejch, and Peter Bell. 2023. The Edinburgh International Accents of English Corpus: Towards the Democratization of English ASR. InICASSP 2023

    work page 2023

  26. [26]

    Shree Harsha Bokkahalli Satish, Gustav Eje Henter, and Éva Székely. [n. d.]. Hear Me Out: Interactive evaluation and bias discovery platform for speech-to-speech conversational AI. ([n. d.])

    work page

  27. [27]

    Berrak Sisman, Junichi Yamagishi, Simon King, and Haizhou Li. 2020. An overview of voice conversion and its challenges: From statistical modeling to deep learning.IEEE/ACM Transactions on Audio, Speech, and Language Processing 29 (2020), 132–157

    work page 2020

  28. [28]

    Zhi Rui Tam and Yun-Nung Chen. 2025. MedVoiceBias: A Controlled Study of Audio LLM Behavior in Clinical Decision-Making.arXiv preprint arXiv:2511.06592 (2025)

    work page arXiv 2025

  29. [29]

    Toyota. 2022. Toyota’s New In-House Intelligent Assistant Learns Voice Com- mands and Gets Smarter Over Time. https://pressroom.toyota.com/toyotas-new- in-house-intelligent-assistant-learns-voice-commands-and-gets-smarter- over-time/

    work page 2022

  30. [30]

    Hao Wang, Luxi He, Rui Gao, and Flavio Calmon. 2023. Aleatoric and epistemic discrimination: Fundamental limits of fairness interventions.Advances in Neural Information Processing Systems36 (2023), 27040–27062

    work page 2023

  31. [31]

    Yihao Wu, Tianrui Wang, Yizhou Peng, Yi-Wen Chao, Xuyi Zhuang, Xinsheng Wang, Shunshun Yin, and Ziyang Ma. 2025. Evaluating Bias in Spoken Dia- logue LLMs for Real-World Decisions and Recommendations.arXiv preprint arXiv:2510.02352(2025)

    work page arXiv 2025

  32. [32]

    Jiao Sun Zhou, Lucie Flek, and David Jurgens. 2024. AI’s language barrier: Coun- tering dialect prejudice in large language models.arXiv preprint arXiv:2405.10846 (2024)

    work page arXiv 2024