"It Would Make a Geordie Feel at Home": Regional Accent Perceptions and Speech Accommodation in Voice AI

1. Introduction

• Recognition accuracy is often lower for speakers whose dialects differ from the linguistic mainstream represented in ASR training data.The introduction frames this disparity as both an access and a sociolinguistic issue.Source: §1 · p. 2
• When technologies fail to recognise a variety, users may modify pronunciation, lexical choices, or accent, reflecting awareness of standard speech and its social meanings.The paper treats voice-AI interaction as a setting for studying accommodation, style-shifting, and language ideologies.Source: §1 · p. 3
• The study examines Geordie speakers’ evaluations of Geordie and SSBE VA voices alongside their reported accommodation and attitudes toward Newcastle and their own accent.Participants interacted with a VA using either Newcastle English or SSBE and took part in post-interaction interviews.Source: §1 · p. 3
• The paper frames VA bias in two directions: users adapt to perceived recognition preferences, while synthetic regional voices are interpreted through established accent stereotypes.It asks how Geordies evaluate regional and mainstream VA voices and how they report accommodating their speech.Source: §1 · pp. 3–4

2.1. Sociolinguistic Biases in ASR

• Dialectal variation is a consistent source of ASR bias: underrepresented varieties tend to fare worse, and disparities may persist despite expanded or adapted training data.The literature also links recognition disparities to systematic linguistic variation, not only data scarcity.Source: §2.1 · p. 5

2.2. Accommodation in Human–Computer Interaction

• Communication Accommodation Theory frames speech adjustment as strategic convergence or divergence; VA users may instead adapt to perceived recognition capacity, an under-studied process for regional dialects.Prior HCI research reports slower, louder, or more careful speech in response to perceived intelligibility problems.Source: §2.2 · pp. 5–6

2.3. Geordie Accent Perceptions

• Newcastle English’s place-linked distinctiveness coexists with prestige stigma, positive social-attractiveness evaluations, and insider–outsider differences; accent can shape perceived competence and credibility.Local perceptions also describe Newcastle speech as broad, fast, melodic, or relatively harsh.Source: §2.3 · pp. 7–8

3.1. Data Collection

• Participants in Newcastle interacted with a simulated VA using either a Geordie or SSBE voice, then completed a post-interaction interview; analyses used interview data rather than experimental speech.The system was controlled by the first author, and its responses were prerecorded by a bidialectal voice actor.Source: §3.1 · pp. 8–9

3.2. Interview coding and qualitative analysis

• Interview responses were descriptively coded and grouped into recurring categories, with verbatim excerpts used to contextualize patterns; only explicitly reported responses were counted.One author conducted the coding, and intercoder reliability was not calculated.Source: §3.2 · pp. 9–10

3.3. Quantitative analysis of accommodation strategies

• Six reported strategies were modeled as repeated binary observations in a mixed-effects logistic regression; demographic predictors did not improve fit, so the strategy-only model was retained.The analysis retained all 57 participants (342 strategy-level observations); estimated probabilities and pairwise comparisons used marginal means and Tukey adjustment.Source: §3.3 · pp. 10–11
• Strategy type significantly improved model fit, indicating that the six accommodation strategies differed in their probability of being reported.The likelihood-ratio test yielded χ2(5) = 28.04, p < .001.Source: §3.3 · p. 11

4.1. Newcastle English Perceptions

• Geordie was associated with both friendliness and warmth and stereotypes of lower intelligence, class, and limited professionalism; participants also described its media representation as underrepresented in authoritative settings and sometimes stereotypical.Some participants recounted accent-based judgements about education and employability, while others rejected the idea that an accent determines a speaker’s abilities or personality.Source: §4.1 · pp. 12–14

4.2. When Do Geordies Feel the Need to Modify Their Accent?

• Professional settings were the most frequently reported context for accent modification (18 participants; 31.6%), followed by travel (14; 24.6%) and family and friends (12; 21.1%).Some participants reported using a broader Geordie accent with family and friends (12 participants; 21.1%). Workplace shifts could aim to aid communication or sound more professional, while travel adjustments included speaking more slowly.Source: §4.2 · p. 15
• Participants described accommodation as often automatic convergence toward an interlocutor or a more formal style, but some reported accent divergence or refusing to change despite recognizing that their accent was socially noticed.Two participants reported that their Geordie became stronger around non-Geordies; four (7%) said they would not change their accent under any circumstances.Source: §4.2 · pp. 15–16

4.3.1. Characterisations of Geordie VA

• Participants most often described the Geordie VA as Geordie (29.8%), familiar (24.6%), friendly (17.5%), and human/natural or having a personality (15.8%).Descriptions primarily emphasized regional and socially engaging qualities rather than function; familiarity could make the voice feel relatable and comforting.Source: §4.3.1 · pp. 17–18, Figure 2 · p. 18
• Some participants found the Geordie VA warm and conversational but not professional, while others viewed its accent as stereotypical or performed rather than authentic.Five of the 35 participants who interacted with the system (14.3%) mentioned stereotypical qualities; these evaluations arose despite the voice being recorded by a native Newcastle English speaker.Source: §4.3.1 · pp. 18–19

4.3.2. Characterisations of SSBE VA

• The SSBE VA was most often described as not a Geordie (22.8%), neutral/normal/generic/robotic (15.8%), or clear/well-spoken (15.8%); overall, it was relatively neutral and clear but less personally familiar than the Geordie voice.Participants had been told they were testing a Geordie VA, making the absence of a Geordie accent salient in the SSBE condition; some associated its neutrality with professionalism.Source: §4.3.2 · pp. 19–21

4.4. Geordie VA Perceptions

• Most participants (n=54, 94.7%) thought a Geordie-speaking VA could be perceived as less authoritative, credible, or reliable than one using a mainstream accent, especially for medical advice.Others said local familiarity could provide comfort, reassurance, naturalness, or intelligibility; one participant found the Geordie voice easier to understand than Southern English.Source: §4.4 · p. 21
• Participants recognized these judgements as socially learned and potentially internalized, while noting that a human-like VA voice could transfer social evaluations onto AI and shape perceptions of suitable advice.Some likened the Geordie VA to an auntie giving informal advice, and participants anticipated more positive evaluations in the North East than in London or elsewhere in the South.Source: §4.4 · pp. 21–23
• Participants linked voice evaluations to accent and gender expectations: female voices seemed gentle and helpful, male voices more forceful or authoritative, and local lexical additions risked sounding forced.They associated Geordie VA appeal with a recognisably regional voice evoking familiarity, belonging, and inclusion, rather than reproducing local features in an exaggerated way; perceived authority could not be attributed to accent alone.Source: §4.4 · pp. 24–25
• Only three participants (5.3%) explicitly rejected the expectation of less favourable evaluations of a Geordie VA; anticipated stigma was located mainly outside Newcastle.Some worried that people from elsewhere might not understand Geordie, while others saw exposure to the accent as welcoming for visitors and potentially increasing acceptance through greater representation.Source: §4.4 · pp. 23–24

4.5. Strategies to Be Better Understood by Voice AI

• 18 participants (31.6%) reported standardising their accent, compared with 16 (28.1%) enunciating and 14 (24.6%) speaking more slowly.Strategy type improved model fit (χ2(5) = 28.04, p < .001); the model-estimated probability was highest for standardising speech or accent (26.0%, 95% CI [14.2, 42.8]). Speaking louder was significantly less likely than enunciating, slowing down, or standardising; avoiding slang was significantly less likely than standardising.Source: §4.5 · pp. 25–26, Figure 4 · p. 26

4.5.1. Strategy 1: Standardise Speech/Accent

• Participants moved away from recognisably Geordie speech toward forms they thought voice AI would recognise, including American, Southern English, “posh,” or more neutral speech.Some described consciously monitoring pronunciation, including clearer /t/s; one participant said he could not change his speech enough to be understood.Source: §4.5.1 · p. 27

4.5.2. Strategy 2: Enunciate

• Participants sometimes suppressed locally characteristic pronunciations through careful articulation, blurring the boundary between enunciation and accent standardisation; they also often combined several accommodation strategies rather than treating them as distinct adjustments.Source: §4.5.2 · p. 28

4.5.3. Strategy 3: Speak More Slowly

• 14 participants (24.6%) reported speaking more slowly, often because they perceived Newcastle English as fast and VAs as having processing limitations.Slowing down could accompany speech modification familiar from interactions with non-Geordies, yet clear and slow speech did not always prevent transcription errors.Source: §4.5.3 · pp. 28–29

4.5.4. Strategy 4: Rephrase/Repeat

• Eight participants (14.0%) reported repeating or rephrasing requests after recognition failures, but some gave up and turned the VA off when attempts continued to fail.These accounts show that accommodation also involved deciding how much effort to invest before disengaging.Source: §4.5.4 · p. 29

4.5.5. Strategy 5: Avoid Slang and Local Lexical Items

• Five participants (8.8%) reported avoiding slang or local lexical items in favour of words they believed VAs were more likely to recognise.Changing a dialect word could also involve rephrasing and modifying accent strength, reflecting expectations that regional forms were less technologically intelligible.Source: §4.5.5 · p. 29

5.1. Accommodation to Perceived Intelligibility

• When recognition failed, participants reported adapting through slower, more careful, or reformulated speech and by avoiding local vocabulary or shifting toward mainstream forms.They described drawing on accommodation practices used with non-local speakers and in professional or institutional settings; these adjustments place interactional effort on users and may eventually lead them to give up.Source: §5.1 · pp. 31–32

5.2. Regional VA Voices: Familiarity vs. Authority

• Participants associated the Geordie VA with familiarity, warmth, personality, and local belonging, while the SSBE voice was more often described as neutral and clear.The Geordie voice could also be seen as less professional, credible, or authoritative; strongly localised features sometimes sounded stereotypical or performative rather than authentic.Source: §5.2 · pp. 32–33

5.3. Convergence and Divergence in Human–Computer Accommodation

• Accommodation can move toward dialect as well as away from it: a Geordie VA prompted one participant to feel more at home and consider speaking more broadly.The discussion distinguishes movement prompted by recognition difficulty from movement prompted by social alignment with the identity represented by the technology.Source: §5.3 · p. 33

5.4. Representation or Being Understood?

• Participants valued regional visibility, but did not uniformly want a VA to reproduce their accent; being understood mattered more for comfortable communication.The discussion argues that regional inclusion should reflect speakers’ own priorities and that recognition of habitual regional speech should not require suppressing dialect features; representation and recognition are both desirable.Source: §5.4 · pp. 33–34

5.5. Limitations and Future Research

• Accommodation findings rely on participants’ metalinguistic reports, not direct speech measurements, so they cannot establish whether reported strategies correspond to measurable production changes.Enunciation and accent standardisation also overlapped in participants’ accounts; because each voice condition used one speaker, accent may be confounded with other acoustic properties.Source: §5.5 · pp. 34–35

6. Conclusion

• Newcastle English speakers reported modifying speech toward forms they considered more standard or intelligible, while Geordie and SSBE VA evaluations reflected established sociolinguistic attitudes.The Geordie voice was linked to warmth and local belonging but also concerns about authority and professionalism; SSBE was seen as more neutral, clear, and credible.Source: §6 · p. 35
• Linguistic inclusion cannot be reduced to increasing the range of accents represented; it also requires reducing recognition disparities and understanding speakers’ preferences.Those preferences are informed by wider sociolinguistic attitudes.Source: §6 · p. 35
• Accommodation, language attitudes, and linguistic stereotyping persist with artificial interlocutors, making sociolinguistics relevant to technology design decisions.The conclusion positions sociolinguistics as a bridge for bringing users’ experiences and linguistic priorities into choices about inclusive voice AI.Source: §6 · p. 36

8. Author contributions: CRediT

• DS led data curation, formal analysis, investigation, methodology, and visualisation.DS also contributed to conceptualisation, funding acquisition, software, and writing.Source: §8 · p. 36
• KT contributed to conceptualisation, funding acquisition, resources, software, visualisation, and writing.KT shared software, visualisation, and writing contributions with the other listed authors.Source: §8 · p. 36
• JS contributed to conceptualisation, funding acquisition, resources, software, visualisation, and writing.JS shared software, visualisation, and writing contributions with the other listed authors.Source: §8 · p. 36

9. Data availability

• The qualitative data are not publicly available because of the need to protect participant anonymity.Anonymised data may be made available on reasonable request through the first author.Source: §9 · p. 36
• The R script and data used to analyse reported accommodation strategies are available on Zenodo.The passage provides the DOI 10.5281/zenodo.23685953.Source: §9 · p. 36
• The authors report using ChatGPT for language refinement and readability enhancement, then reviewing and editing the content and taking responsibility for it.The disclosure appears under the manuscript-preparation declaration.Source: §10 · pp. 36–37

• The participant distribution is organized by age, gender, and VA system; the Geordie VA condition had more participants than the SSBE comparison condition.Source: Appendix A · p. 37, Table A.1 · p. 37

• Participants completed a semi-structured interview after the experimental interaction.Source: Appendix B · p. 37

• Participants were asked what they thought the experiment was about.Source: Appendix B.1 · p. 37

• Participants were asked to rate response accuracy and whether the system understood their accent and dialect without significant issues.Source: Appendix B.2 · p. 37

Appendix B.3. Interaction Comfort and Naturalness

• Participants were asked whether they felt comfortable speaking with the VA and how natural the interaction felt.Source: Appendix B.3 · p. 38

Appendix B.4. Speech Adaptation Strategies

• Questions explored whether participants altered their speech during the interaction and what strategies they used to be better understood.Source: Appendix B.4 · p. 38

Appendix B.5. Perceived System Bias and Inclusivity

• Participants were asked whether they perceived dialect bias in the VA and how inclusive it seemed for speakers of their dialect.Source: Appendix B.5 · p. 38

Appendix B.6. Emotional and Behavioral Responses

• The guide asked about feelings during interaction, preference for dialect-sensitive or standard systems, and descriptions of the VA.Source: Appendix B.6 · p. 38

Appendix B.7. Suggestions for Improvement

• Participants were asked what improvements they would suggest and where the system failed to meet expectations.Source: Appendix B.7 · p. 38

Appendix B.8–B.9. Previous Experience and Background

• Questions covered general VA experience and usage, speech adaptation with technology, technology attitudes, and demographic and background information.Source: Appendix B.8 · p. 38, Appendix B.9 · p. 39

Appendix B.10. Regional Orientation and Community

• The guide asked about connection to the community where participants’ dialect is spoken, their feelings about Newcastle, and thoughts of moving away.Source: Appendix B.10 · p. 39

Appendix B.11. Dialect, Identity, and Representation

• Questions addressed dialect identity and social use, representation in media and technology, desired technology support, Newcastle features and examples, and VA omissions.Source: Appendix B.11 · p. 39