By
Stuart Kerr, Technology Correspondent,
LiveAIWire

Ask an AI speech recognition system to transcribe a conversation
in received pronunciation and it will likely perform with high accuracy. Ask
it to transcribe the same content spoken in a Glaswegian, Geordie, or Black
Country accent and error rates rise substantially. Ask it to handle the
slang, code-switching, and dialectal vocabulary of regional British
vernacular and performance degrades further still. This is not a minor
calibration issue. It is a structural consequence of how large speech and
language models are built, and it produces real inequity in who benefits from
AI-assisted services and who is disadvantaged by them.

Speech recognition systems are trained on datasets assembled to
maximise coverage and quality. In practice, maximising quality has
historically meant overweighting recordings of speakers with accents
associated with broadcast media, academic institutions, and corporate
environments: received pronunciation in the UK, general American in the US.
Research
from the University of Edinburgh on accent bias in AI has
documented systematic performance gaps across UK regional accents, finding
error rate differentials of up to 20 percentage points between speakers of
standard southern English and speakers of some northern and Celtic regional
accents. The gap is not explained by acoustic complexity alone. It reflects
training data composition that systematically underrepresents regional
speech.

Why Slang Is Harder Than Accent

Accent presents a phonological challenge: mapping unfamiliar sound
patterns to known words. Regional slang presents a lexical and semantic
challenge that is structurally more difficult. Terms that are standard in one
regional community are either absent from AI training corpora or present in
quantities too small to establish reliable semantic associations. When a
language model encounters low-frequency regional vocabulary, it defaults to
the nearest high-frequency equivalent, which may have an entirely different
meaning, or flags the term as unknown.

The problem compounds with code-switching, the common practice
among bilingual and bidialectal speakers of moving between registers within a
single utterance. Welsh-English speakers regularly blend Welsh grammatical
constructions into English sentences. British South Asian speakers deploy
Punjabi, Urdu, or Gujarati terms within English-dominant speech in ways that
carry cultural specificity that disappears entirely in AI transcription and
translation. AI systems trained primarily on monolingual standard English
corpora have no framework for handling these switches, producing
transcriptions that either drop the code-switched element or replace it with
a phonetically similar English term that changes the meaning of what was
said.

The Practical Consequences

The failure of AI systems to handle regional British speech
affects people unequally in proportion to how far their speech diverges from
the training data standard. AI-assisted services including voice-controlled
customer service systems, healthcare appointment scheduling, legal
transcription tools, and accessibility aids for deaf and hard-of-hearing
users all deliver worse outcomes to people whose speech reflects regional,
ethnic, or socioeconomic backgrounds underrepresented in AI training data.
The irony is that these are frequently the populations with the greatest need
for the accessibility benefits that AI voice interfaces
promise.

As our analysis of AI’s
broader challenge with language diversity and dialect found, the
gap between AI language capability in high-resource and low-resource
linguistic contexts is a governance and investment question as much as a
technical one. The BBC
Research and Development programme on accent diversity in voice
technology has been actively documenting and addressing accent bias
in speech systems since 2019, finding consistent evidence of performance gaps
across regional British accents and developing datasets intended to improve
coverage. That work is valuable, but the resources dedicated to expanding
training data for regional British speech are modest relative to the
commercial incentive to optimise for the largest user base, which remains
speakers of standardised varieties.

What Better Looks Like

The technical path to reducing accent and dialect bias in AI
speech systems is understood even where it has not been adequately resourced.
Expanding training datasets to include systematically collected speech from
regional speakers, in collaboration with communities rather than through
passive data harvesting, improves baseline performance. Fine-tuning deployed
models on regional speech data allows adaptation without full retraining.
Explicit evaluation of model performance across accent and dialect
categories, included as a standard deployment requirement rather than an
optional post-hoc audit, creates incentives for developers to address gaps
before deployment rather than after complaint.

The governance requirement is equally important. Procurement
frameworks for AI voice and language systems used in public services should
specify minimum performance standards across regional accent categories as a
contract condition, rather than accepting the default performance of
commercial systems optimised for majority user bases. Without that
requirement, the market incentive to serve regional speakers adequately does
not exist. As our coverage of how
AI systems encode and perpetuate social inequalities through training data
bias found, the distributional effects of AI performance gaps are
not neutral. They fall most heavily on communities whose voices are already
least represented in the institutions and systems that affect their lives,
and they require structural governance responses rather than voluntary
improvement commitments.

The Healthcare and Justice Implications

The consequences of accent and dialect bias in AI speech systems
extend beyond inconvenience in consumer applications into domains where the
failure to understand someone accurately has serious consequences. AI
transcription tools used in police interviews, court proceedings, and
healthcare consultations are being deployed in settings where an inaccurate
transcript can affect legal outcomes or medical treatment. A patient from a
regional background whose speech the AI clinical documentation system
misrenders may have errors introduced into their medical record that persist
and compound over subsequent encounters with the health system. A defendant
whose interview transcript contains systematic errors introduced by
accent-biased speech recognition may face a criminal record built on flawed
documentation.

These are not hypothetical risks. Researchers in linguistics and
legal studies have documented cases where AI speech recognition errors in
legal and clinical settings have produced documentation that diverged
materially from what was said, in ways that native speakers of regional
varieties would immediately recognise but that standard-English-trained
reviewers did not catch. The governance response required here is specific
and achievable: mandatory accent diversity testing for AI speech systems
deployed in legal and clinical settings, as a procurement condition rather
than a post-deployment audit. As our coverage of how
AI systems perform differently across populations in high-stakes
contexts found, the populations most harmed by AI performance gaps
are consistently those with the least power to contest the decisions those
gaps produce. Regional speakers navigating legal and healthcare systems are
among the clearest cases of that general observation, and they deserve the
same standard of accuracy that speakers of standardised varieties receive as
a baseline expectation rather than an aspirational goal.

What Should Change

The technical path to reducing accent and dialect bias in AI
speech systems is understood even if it is not being pursued at the pace that
the equity implications demand. Assembling training datasets that represent
regional British speech in proportion to its prevalence in the population,
rather than in proportion to its presence in easily scraped digital audio,
requires deliberate effort and investment. The BBC Listening Project and
similar archival initiatives have created resources that, combined with
purpose-built recording programmes targeting underrepresented accents, could
substantially improve performance on regional varieties within a realistic
development timeline.

The commercial incentives do not naturally align with this
investment. Companies developing AI speech systems prioritise performance for
the largest and most commercially valuable user groups, which creates a
self-reinforcing dynamic in which underrepresented accents remain
underrepresented because the market for tools that serve those accents is not
as large as the market for tools that serve dominant ones. Addressing this
requires either regulatory intervention that establishes performance
standards across linguistic variety, or public investment in developing and
sharing diverse training datasets that commercial developers can use without
bearing the full cost of their creation.

For related coverage, see our analysis of whether
AI is erasing linguistic diversity and our broader look at why
mitigating AI bias is harder than it looks.

About the Author

Stuart Kerr is the Technology Correspondent for LiveAIWire. He
writes about artificial intelligence, emerging technology, and the forces
reshaping work, business, and society.