All articles
Guide9 min read

Can AI Handle Accents, Patient Names and Medication Names?

Yes, an AI receptionist can understand most accents, patient names and medication names. A safe one confirms anything important and passes uncertain details to a person instead of guessing.

Clinics in Aotearoa ask us this question a lot. A patient rings from Tāmaki Makaurau with a strong accent and spells out a te reo Māori surname. Then they ask for a repeat of a medicine whose name sounds like three others. Can the AI keep up?

Speech recognition has improved a great deal, but raw accuracy is only part of the answer. In a clinic, what counts is whether the system was built for clinic calls, what it does when it isn’t sure, and who checks the result before it reaches a clinician.

We build Heron, the AI operating layer for clinics, so we have a clear view on this.

Can an AI receptionist understand accents, patient names and medication names?

Mostly, yes, if it was built for clinic calls. Current speech recognition handles a wide range of accents and everyday clinic vocabulary well. Unusual names and medicine names are harder, and that is where generic AI receptionists built on off-the-shelf tools slip. A good AI receptionist checks those details against the clinic’s records and hands anything uncertain to your team. It does not fill the gap with a guess.

We care most about that second part. A clinic does not need a system that is usually right and says nothing when it gets something wrong. A clinic needs a system that knows when it is unsure.

Speech recognition in 2026: what has improved

Speech-to-text models now train on huge amounts of multilingual audio. That training gives them a much better ear for accents than the old phone menus had, where you had to speak slowly and stick to a narrow set of words.

A modern model copes well with everyday requests. Booking an appointment, moving a time and asking whether the nurse does immunisations on Saturday all use common words, and the context helps the model.

Where speech recognition accuracy still drops

Accuracy falls on the words that matter most in healthcare: names, medicines and medical terms. These words are rare in training data, and many of them sound like other words.

The research shows the same pattern. A 2024 study tested 19 speech recognition models on clinical English spoken with 93 African accents. Whisper-large, one of the stronger models tested, recalled only 42% of medication names and 33% of protected health information, such as names and ID numbers. A clinic cannot afford to get those details wrong.

How an AI receptionist copes with accents on a phone line

Phone audio makes accents harder to recognise. Calls come through compressed, sometimes from a car or a noisy kitchen, and the caller is often unwell. A model that handles an accent well in a quiet room has a harder job on a crackly mobile line.

Some vendors let you choose the agent’s own accent. Lyngo, for example, offers local-accent voices. A familiar voice helps patients understand the agent, but it doesn’t change what the agent hears. Recognising what the caller says is the harder half of the conversation.

Why generic AI receptionists get accents and names wrong

Many AI receptionists are built on off-the-shelf voice agent tools: a general-purpose speech model, a stock voice and a script on top. That gets a phone answered quickly, and it copes with simple requests. What it can’t do is change how the model hears or speaks. Every caller gets the same model, trained mostly on American and British English, whether they’re saying a te reo Māori surname, a local street name or a medicine name.

That is where generic agents fall down, on the words a clinic can least afford to get wrong. They mishear unfamiliar names and healthcare terms, mispronounce names when they read them back, and have no way to learn your clinic’s own vocabulary. The research above shows the gap: a general-purpose model can do well on everyday speech and still miss more than half of medication names.

Heron doesn’t rent its voice from an off-the-shelf tool. It runs on Parrot, our own voice pipeline built for clinics, so we control how Heron hears and how it speaks:

A generic agent can sound polished in a demo. Ring it with your own patient list and the difference shows.

How an AI receptionist should handle uncertainty

Whether an AI receptionist is safe depends less on the model and more on the checks built around it. Here is the pattern we think every clinic should expect from any vendor:

  1. Verify identity against your records. The agent checks the patient’s details against the PMS before it acts. Heron verifies the patient from their name, date of birth and phone. When someone calls on a patient’s behalf, Heron verifies both the caller and the patient.
  2. Ask again when a detail doesn’t match. A name that matches an enrolled patient is more reliable than free text, so a mismatch should mean another question, not a guess.
  3. Raise what it can’t confirm. When the system isn’t sure about a detail, a person checks it. Anything Heron can’t confirm, it raises to your team.
  4. Hand over cleanly. When the caller wants a person, or the request needs one, your team gets the request, the context and a summary.

Patient names and cultural respect

Getting a patient’s name right is basic respect. Patients notice when a clinic gets their name wrong. A wrong name is also a safety problem, because it can attach a request to the wrong record.

When an AI phone agent mispronounces a name

Hearing a name and saying a name are separate problems. An agent can mishear a name. It can also capture a name correctly and then say it badly. Patients find both careless.

Checking against the patient record fixes mishearing. Pronunciation needs its own fix, which is why Parrot lets clinics set custom pronunciations. Even so, the spoken name should never be the only check. When the patient’s name, date of birth and phone match the record, the booking is right even if the agent’s pronunciation is a little off. In our experience, patients forgive an awkward pronunciation far more easily than a booking under the wrong name.

Te reo Māori names

Clinics in Aotearoa hear te reo Māori names every day. Those names have vowel sounds and patterns that speech models trained mostly on American and British English were not built around. Names like Ngaio, Whetū and Tūī deserve to be heard and said correctly.

This is where a generic agent struggles most, and where a pipeline built for Aotearoa matters most. Whichever vendor you choose, test it with real local names before it goes live. Ring it yourself. Use names like the ones on your own patient list, including macrons and spellings that don’t follow English rules. Listen to what it captures and how it says them.

Verify before anything reaches the clinical record

Heron isn’t your clinical record. Clinical records, diagnoses, treatment notes and medical history stay in your PMS. Heron reads what it needs to verify a patient and do the work, and writes bookings, requests and tasks back to the PMS for your team.

AI voice agents and medication names

Medication names carry the highest risk. Many medicine names sound alike, the dose matters, and a caller reading a label over the phone won’t always say it clearly.

That is why we treat medicines as a handoff. Heron captures the repeat prescription request, does the pre-work and routes it to the clinician responsible for approval. Heron never signs or approves a prescription. The clinician makes the decision, with the request already organised for them.

The pre-work is where look-alike names get caught. In Medtech, Heron checks the patient’s current medications and confirms which item they mean, so the request matches a medicine already on the patient’s record, not a word heard on a phone line. The request then becomes a task for the clinician, or a named staff member, marked urgent or normal, or a note in the daily record. The clinician prescribes from Medtech as usual, with the usual interaction and allergy checking. See how Heron handles repeat prescriptions in Medtech.

Healthcare compliance: where the call data goes

Voice AI handles sensitive health information, so ask every vendor where that information goes. Heron’s primary database and core data storage are in Australia. Some infrastructure vendors are in the United States: providers for API hosting, telephony, AI processing such as real-time speech-to-text and text-to-speech, and payment processing. A complete, current list of overseas recipients is available on request.

Heron’s SOC 2 Type I was independently audited in March 2026. Data is encrypted at rest using Google Cloud KMS and in transit using TLS 1.2+. Access is role-based, with mandatory MFA through SSO. Heron is designed to comply with the New Zealand and Australian privacy acts and the Health Information Privacy Code, and is HIPAA aligned. See security at Heron.

Your clinic keeps responsibility for record retention, because your PMS holds the permanent record. In New Zealand, the 10-year requirement under the Health (Retention of Health Information) Regulations 1996 sits with the clinic.

What Tamatea Medical Centre saw

Tamatea Medical Centre, a high-volume general practice in Hawke’s Bay, runs Heron alongside its reception team. The clinic named its Heron agent Scout and introduced it to patients as the clinic’s new trainee, which helped staff and patients adopt it.

At Tamatea Medical Centre, average hold times are now under 5 minutes, down from 16 to 20 minutes, and 66% of patients self-serve through Heron with zero wait. Anyone can opt out to a person.

That choice matters here. Speech recognition will never suit every caller, and it doesn’t have to. Patients who want a person get a person, and the reception team handles the conversations that need a human. Read Tamatea’s story.

What to test before you choose

Ring it yourself with names, accents and look-alike medicines from your own patient list, then ask for a person mid-call and see what your team receives. To try this on Heron, book a demo.

Sources

  1. Afonja T, Olatunji T, Ogun S, et al. Performant ASR models for medical entities in accented speech. Interspeech, 2024.
  2. Lyngo: lyngo.ai
  3. Heron: Security, Changelog, Tamatea Medical Centre customer story
Keep reading

More from Heron

All articles

Ready to see how Heron can help your clinic?

Book a demo and see Heron answer your patients, book straight into your PMS, and hand your team only what needs a human.