Summary
Speech recognition is measurably worse on British accents than American ones, because the training data is not evenly distributed. Published benchmarks find the lowest error rates on US accents, and Scottish, Welsh and Northern English speech among the most often misrecognised. It is a real effect and a testable one. Record ten minutes of your own clinic, with your own voices, and read the examination section first.
Almost every ambient scribe sold into the UK was trained predominantly on American English, and speech recognition accuracy follows the training data. Benchmarks consistently find the lowest word error rates on the accents best represented in the corpora, which means American and Canadian English first, with British, Scottish, Welsh and Northern English speech faring worse.
This is not a reason to avoid the category and it is a reason to test differently. Vendor accuracy figures are almost always quoted from evaluations that do not tell you which accents were in the sample, and a 98 percent claim measured on Californian speech says very little about a Tuesday clinic in Dundee.
The good news is that it is one of the few product claims a clinic can check in an afternoon, on its own voices.
See it in your clinicTest it on your own list before you believe anyone's accuracy number, including ours.
Request a demo| Test | Why it is the one that matters |
|---|---|
| Your own clinicians, not a demo voice | Accuracy follows training data, and the demo is chosen to flatter |
| The examination, spoken aloud | Research puts scribe error highest on the physical examination |
| Every number read back | A wrong figure in a record is the error that reaches a patient |
| Negations | No chest pain becoming chest pain is the classic failure |
| Patients as well as clinicians | The patient's accent is in the audio too, and is rarely tested |
| Drug and condition names | Local and brand terminology is where general models thin out |
| A noisy room | A clinic is not a recording booth, and neither is a demo |
What the benchmarks actually show
Accent bias in automatic speech recognition is well documented and it tracks the training corpus. Comparative work finds the best word error rates on the accents most represented in training data, namely American and Canadian English, with Irish, English and New Zealand accents close behind, and measurable gaps for others.
One evaluation across major commercial engines reported native UK English speakers among the highest word error rates in the sample. Work on Scottish regional accents found that baseline models produced systematically higher error rates without adaptation, and that accent-specific fine tuning reduced them, which is the useful half of the finding because it means the gap is closeable rather than fixed.
Resources like the Edinburgh International Accents of English Corpus exist precisely because the gap is measurable. Northern English, Welsh and Scottish speech are consistently among the more often misrecognised in English language benchmarks.
Why a clinic feels it more than a benchmark does
A benchmark reads prepared sentences in a quiet room. A consultation has two accents in it rather than one, overlapping speech, a door, a corridor, a child, and vocabulary that general models see rarely.
The patient's accent is the half nobody tests. A vendor demo has the clinician speaking, and the clinician is usually the person who chose the product. The patient describing their symptoms is in the audio too, and in most UK clinics that is a wider range of speech than the staff.
Names are the sharpest edge. Local place names in a history, drug brand names, and patient names from every language spoken in the catchment. A model that handles a Home Counties consultation cleanly can still mangle a name that matters.
The habit that closes most of the gap
Speak the findings you want recorded. This sounds like a workaround and it is closer to a discipline worth having anyway, because a scribe records what was said and infers what was not, and inference is where the error rate lives.
Simulated-encounter research puts scribe error highest on the physical examination, which is exactly the section clinicians perform in silence. A clinician who narrates the examination gets a better note from any product, and the improvement is larger than the difference between most products.
Read numbers back aloud. In cardiology, ophthalmology and anywhere else the record is measurements, that one habit removes the failure mode that actually reaches patients. Our review of the hallucination evidence covers where the errors concentrate.
How to test it in an afternoon
Take ten minutes of real consultation audio with consent, choose the clinician whose accent is strongest rather than the one who demos well, and include a patient rather than a colleague reading a script. Run it through every product on the shortlist.
Then read the output in this order. The examination section first, because that is where inference happens. Every number against the source. Every negation. Then the names.
Score it as a count rather than an impression. Errors per hundred words, and separately, errors that would have reached the patient. Those are different numbers and only the second one decides anything. The wider evaluation design is in running a scribe pilot.
What to ask a vendor
Ask which accents were in the evaluation behind their accuracy figure. Most cannot answer, and a vendor who says so plainly is more useful than one who produces a number without a denominator.
Ask whether the model was adapted for British English or is a general model with a UK interface. Ask what happens to a word the model is unsure of, because a product that flags uncertainty is safer than one that guesses fluently.
Ask what you can do about it after purchase. Custom vocabulary, clinician-specific adaptation and a correction loop that actually improves future output are all real features, and they are the difference between a fixable gap and a permanent one. The full list is in what to ask a scribe vendor.
Where wai sits
wai is built inside a working London clinic, so the audio it was shaped on is British, mixed, and recorded in rooms rather than booths. That is a reason to test it rather than a claim to believe.
It shows its working, so a reviewing clinician can see what came from the audio and what was inferred, which is the control that matters most when the recognition is uncertain.
Bring your hardest ten minutes. If another product reads your clinic better, buy that one. Clinicians still decide what enters the record.
Bring your hardest ten minutes of audio. Fifteen minutes, on your own voices.
Request a demoFrequently asked questions
Do AI scribes struggle with British accents?
Measurably more than with American ones. Speech recognition accuracy follows the training corpus, and benchmarks consistently find the lowest word error rates on American and Canadian English. Scottish, Welsh and Northern English speech are among the more frequently misrecognised in English language evaluations.
Which UK accents are hardest for speech recognition?
Scottish, Welsh and Northern English speech appear most often in the literature on recognition gaps. Work on Scottish regional accents found baseline models produced systematically higher error rates without adaptation, and that accent-specific fine tuning reduced them, so the gap is closeable rather than fixed.
Can I trust a vendor's accuracy percentage?
Not without asking which accents were in the evaluation. A 98 percent figure measured on American speech tells you very little about a clinic in Dundee or Swansea. Most vendors cannot answer the question, and one that says so plainly is more useful than one producing a number with no denominator.
How do I test a scribe on my own accents?
Ten minutes of real consultation audio with consent, using the clinician whose accent is strongest rather than the one who demos well, and including a patient rather than a colleague. Run it through every shortlisted product, then read the examination section first, every number, every negation and then the names.
Does the patient's accent matter too?
Yes, and it is the half nobody tests. A vendor demo has the clinician speaking, and the clinician usually chose the product. The patient describing symptoms is in the audio as well, and in most UK clinics that is a wider range of speech than the staff.
What single habit most improves accuracy?
Speaking the findings you want recorded, especially during the examination. Research puts scribe error highest on the physical examination, which is the section most clinicians perform in silence. Narrating it improves the note from any product by more than the difference between most products.
Can anything be done after purchase?
Often yes. Custom vocabulary, clinician-specific adaptation and a correction loop that improves future output are real features worth asking about. They are the difference between a gap you can close and one you live with for the length of the contract.
Sources
- English accent accuracy analysis in a state-of-the-art automatic speech recognition system
- Evaluating OpenAI's Whisper ASR, performance analysis across diverse accents and speaker traits, JASA Express Letters
- Human and computer recognition of regional accents and ethnic groups from British English speech
Test it on your own voices
Fifteen minutes on your workflows. If it is a fit we reply within 24 hours.
Partner with us
See it on your own patients, then build it with us
Thirty minutes on your own patients, doctors and rules, in your market. We reply within one working day, and the clinics that partner with us early shape what we build next.
Read how multilingual care works in a clinic