Key takeaways
- Recognition problems fall hardest on the patients who already find access difficult, so test before go-live, not after complaints.
- Build a test set from your own population: local accents, older voices, second-language speakers, poor mobile lines and noisy rooms.
- Score names, dates of birth, numbers and place names separately; that is where most errors happen.
- Every failure should fail safe: a clear route to a person, never a loop.
- Plan for languages other than English and for disabled callers under the Equality Act duty to make reasonable adjustments.
Why this matters for access
Speech recognition has improved enormously, but it is still uneven. It tends to do best with the voices it heard most during development and worst with strong regional accents, older voices, second-language speakers and anyone calling from a noisy kitchen or a bus. In a practice, those are often the callers who already find it hardest to get through.
The access gap is real without any AI involved. Adults renting from a council were 46% more likely, and those renting from another social landlord 51% more likely, to report difficulty contacting their GP practice than outright homeowners[1]. The population is also ageing: in mid-2025 there were 285 people of pensionable age for every 1,000 people of working age[2]. A voice tool that works well for a confident 35-year-old on a good line but struggles with an 85-year-old in Glasgow or Newcastle can quietly widen the gap.
The language picture
In the 2021 Census for England and Wales, 91.1% of usual residents aged three and over had English (or Welsh in Wales) as their main language. A further 1.5%, about 880,000 people, could not speak English well, and 0.3%, about 161,000, could not speak it at all[3]. Those national figures hide very large local differences, so look at your own practice list and local census data.
In Wales, an estimated 538,300 usual residents aged three and over, 17.8% of the population, reported being able to speak Welsh[4]. Practices there should check how a voice tool fits their Welsh language commitments before any patient hears it.
Where recognition usually goes wrong
- Names: unusual spellings, names from other languages and surnames that sound like common words.
- Numbers: fifteen and fifty, dates of birth said in different orders, phone numbers with pauses.
- Place names: local streets, villages and estates that do not appear in national data.
- Medical words said in everyday terms: 'water infection', 'me chest', 'the sugar'.
- Mixed language: a caller switching between English and another language mid-sentence.
- Speech differences: stammering, slurred speech after a stroke, voices affected by Parkinson's, hearing aid users who speak loudly.
- Conditions: speakerphone, wind, traffic, a television in the background, or a weak mobile signal.
Each of these can push a tool into guessing. The test is not whether it ever mishears, because people mishear too, but whether it notices it may have misheard and checks, or routes the caller to a person.
Building a test set from your own patients
Profile your list
Use your practice demographics and local knowledge to list the accents, languages and age groups your callers actually have.
Recruit voices
Ask staff, members of your patient participation group and local volunteers to record test calls. Aim for at least two voices for each group you identified.
Write a script
Use the same twenty utterances for every voice so results compare fairly, using fictional names and dates only.
Vary the conditions
Record each script on a mobile outdoors, on a landline, on speakerphone and with background noise.
Score separately
Mark names, numbers, place names and intent separately, and note whether the tool checked back when it was unsure.
Repeat after changes
Run a shorter version whenever the supplier updates the system or you change the wording.
A test script you can adapt
| Utterance type | Fictional example | What to check |
|---|---|---|
| Opening request | "I'm ringing about getting an appointment for me mam" | Intent recognised; third-party caller handled correctly |
| Name | "It's Siobhan Ní Bhriain" | Spelling checked back, not guessed |
| Date of birth | "Third of the fifth, fifty-two" | Date read back in a clear format |
| Number | "Is it fifteen or fifty minutes?" | Ambiguity resolved |
| Local place | A street or estate near your practice | Recognised or asked to repeat |
| Everyday symptom words | "I've got a water infection again" | Routed to the right process, no triage |
| Urgent words | "He's gone all grey and he can't breathe" | Agreed 999 wording at once |
| Language request | "Polish, please" or "Do you speak Urdu?" | Interpreter or alternative route offered |
| Ask for a person | "Can I just talk to someone?" | Transfer or direct line offered |
Scoring and deciding
Agree pass criteria before testing so the results cannot be reinterpreted afterwards. The safety criteria are absolute: every urgent utterance, in every voice and every condition, must get the agreed urgent-help wording. For everything else, look at both accuracy and behaviour when unsure.
Example (fictional) criteria: intent recognised in at least 19 of 20 utterances for every voice group; no name or date of birth accepted without a read-back; no voice group more than two errors worse than the best group; every failed utterance ends with a route to a person.
The third criterion matters most. An average that looks good can hide one group of patients for whom the tool barely works. If a group fails, ask the supplier what can change, retest, and if it still fails, consider keeping that group's likely calls, such as a dedicated line for a care home or a community clinic, with staff.
Languages other than English
A practice voice tool does not need to speak every language, but it does need a plan for callers who do not speak English well. Options include recognising a language request and routing the caller to a member of staff who can book an interpreter, offering translated information on the website and in texts, and making sure the tool does not keep repeating English prompts to a caller who is clearly struggling.
- Record which languages callers request during the pilot, to plan interpreter cover.
- Make sure staff know how to add an interpreter to a call quickly.
- Keep any translated wording approved by the practice, not generated on the fly.
Disabled callers and reasonable adjustments
Under the Equality Act 2010, service providers must take reasonable steps to avoid substantial disadvantage to disabled people, including by providing auxiliary aids and, where information is involved, providing it in an accessible format[5]. For a phone tool that means thinking about callers with speech impairments, hearing loss, learning disabilities or cognitive impairment, and making sure they can bypass the tool easily.
- Flag patients who have asked for adjustments so their calls can go straight to a person where your system allows.
- Keep a text relay, email or online alternative visible on the website and in the greeting.
- Let callers ask for a person at any point with simple words, and test that it works with slurred or slow speech.
Questions for the supplier
Your own test is the evidence that counts, but the supplier's answers tell you how seriously they take the issue and how much help you will get when something fails. Ask in writing:
- Which accents, age groups and speech differences have you tested with, and can we see the results by group rather than as a single average?
- What does the tool do when its confidence in what it heard is low, and can we adjust that threshold?
- Can callers ask for a person in any words at any point, and how is that tested?
- How do you add local names, streets and estates so they are recognised?
- What happens to recognition when you update the system, and will you tell us before an update goes live?
- Can we export a sample of transcripts to check recognition ourselves?
Vague answers such as 'it works with all accents' are a warning sign. A supplier who has done the work will show you where the tool is weaker and what they do about it.
Example: what a test reveals
Example (fictional): an inner-city practice tests a voice tool with 14 voices, including staff with Glaswegian, Scouse and Nigerian English accents, two older patient volunteers and a Polish-speaking receptionist.
In this example intent recognition is strong across almost all voices, but dates of birth from the two older volunteers on speakerphone are misheard several times, and the tool accepts them without reading them back. The supplier changes the setting so every date is read back. On retest the error rate is the same, but every error is caught. The practice decides that is acceptable for information calls and keeps identity-based tasks with staff for now.
Keep testing after go-live
Monitoring should look for patterns by group, not just overall accuracy. Sample calls from different times of day, ask reception to flag any patient who mentions they struggled with the automated line, and rerun a short version of the test set every quarter and after every significant update. Accent and language performance is never a one-off checkbox; it is part of keeping access fair.
Frequently asked questions
Do AI voice tools struggle with regional accents?
Performance varies by product and by voice. The only reliable way to know is to test the tool with voices and conditions that match your own patients before go-live.
What should happen when the tool mishears a caller?
It should check back, for example by reading a date of birth back, and if it is still unsure it should offer a route to a person rather than guessing or looping.
Does a practice voice tool need to speak other languages?
Not necessarily, but it needs a plan: recognise a language request, route the caller to staff who can arrange an interpreter, and avoid repeating English prompts to someone who is struggling.
How do reasonable adjustments apply to AI phone tools?
The Equality Act duty still applies, so disabled callers must be able to bypass the tool easily and have accessible alternatives such as text relay or online contact.
References
Office for National Statistics (2026). Experiences of NHS healthcare services in England: September 2026 (opens in a new tab) ↑
Office for National Statistics (2025). Population estimates for the UK, England, Wales, Scotland and Northern Ireland: mid-2025 (opens in a new tab) ↑
Office for National Statistics (2022). Language, England and Wales: Census 2021 (opens in a new tab) ↑
Welsh Government (2022). Welsh language in Wales (Census 2021) (opens in a new tab) ↑
legislation.gov.uk (The National Archives) (2010). Equality Act 2010, section 20: duty to make adjustments (opens in a new tab) ↑