A plumber I spoke to earlier this year had switched off his AI receptionist with a Kiwi accent after four days. His reason was fair enough. A customer had said, “I’ll be there at ten past eight,” and the booking came through as 6:10. Another caller in Whangārei got asked three times to repeat the suburb. He decided the technology was not ready for New Zealand.
The technology was fine. The setup was not. Almost every complaint about an AI receptionist traces back to a handful of specific, fixable things: which speech model is doing the listening, whether local names were ever loaded into it, and whether anyone built a confirmation step before the booking got written.
This post covers what actually goes wrong, why our vowels cause it, and how to test a system properly before you put it on your main line.
Key Takeaways
- An AI receptionist mishears a Kiwi accent because of vowel shifts and training data gaps, not because the caller is unclear. The fix sits in the setup.
- New Zealand English raises short front vowels, so “ten” lands as “tin” and “deck” as “dick”, which wrecks numbers, dates and street names first.
- Māori place names and surnames break most off-the-shelf voice models until you load them in as custom vocabulary with the pronunciations locals actually use.
- Many “it misheard me” complaints are turn-taking faults, where the system talks over a caller who paused mid-sentence, not transcription errors at all.
- Confirmation loops beat perfect transcription. Read the number back and spell the surname, and a 92% accurate system behaves like a 99% one.
- Never ask a caller to say an email address out loud. Capture the mobile from caller ID, confirm it, and text them a link instead.
- Test with twenty real calls before go-live, including one from a ute on the motorway with the window down and one from a noisy café.
- Dr Claire’s practice went from 47% of calls unanswered to none missed, with booked appointments up 44%, once the phone was handled properly.
Why an AI Receptionist Mishears the Kiwi Accent
Speech recognition is not listening the way a person listens. It is running audio through a model that has been trained on enormous amounts of recorded speech, then guessing the most probable sequence of words. The guess is only as good as the examples it was trained on.
New Zealand has five million people. In a training set built mostly from North American and British recordings, we are a rounding error. Australian English gets more coverage, and because the two accents look similar from the outside, a lot of models effectively treat us as slightly odd Australians. We are not, and the differences sit exactly where phone calls carry the most important information.
Our vowels moved, and the numbers moved with them
The short front vowels in New Zealand English have shifted upward. The vowel in “dress” has raised toward “kit”, so “ten” comes out closer to “tin” and “deck” closer to “dick”. The vowel in “trap” has raised toward “dress”, so “bag” leans toward “beg”. Meanwhile our “kit” vowel went the other way and centralised into something closer to “uh”, which is why the fish and chips joke exists.
Now put that on a phone line where somebody is giving you a time, a date, or a street number. “Seven” and “eleven” both carry that raised vowel. “Ten” and “tin” are a single misread apart. A model trained on American speech has a strong prior that a person saying something near “tin” means the metal, not the number, because in its training data that is nearly always true.
We are fast, clipped, and we swallow consonants
Kiwis speak quickly and drop a lot of endings. Auckland becomes “Ak-lənd”. “Going to” becomes “gunna”. Final consonants get softened or glottalised. Add the high rising terminal, the upward inflexion at the end of a statement, and you have sentences that sound like questions to a system watching pitch to work out whether the speaker has finished.
There is a real research base under this. A Stanford study published in PNAS found commercial speech recognition systems produced substantially higher word error rates for some speaker groups than others, driven by imbalanced training data rather than anything about the speakers. Accent gaps in voice AI are a known, measured property of these systems. They are not a mystery, and they are not a reason to give up.
Te reo Māori place names are a category of their own
Whangārei, Ngāruawāhia, Papatoetoe, Ōtāhuhu, Taupō, Te Atatū, Waiuku, Manurewa. A model with no exposure to Māori phonology will produce something unrecognisable; then the booking system will fail to match the suburb; then the call falls over. The same applies to surnames. Ngata, Waititi, Te Whata, Kereopa. These are common New Zealand names, and an untuned system will mangle every one of them.
None of this means voice AI cannot answer your phone. It means an out-of-the-box configuration built for Ohio will make you look bad in Ōrewa.
Where the Kiwi Accent Breaks an AI Receptionist First
When I audit a voice system that is underperforming, the failures cluster. It is almost never a general inability to understand the caller. It is five or six specific moments in the call where the wheels come off.
Numbers, times and dates
This is the biggest one, and it is worth understanding why. In ordinary conversation, context rescues a misheard word. If the model gets one word wrong in a sentence about a leaking tap, the meaning survives. Numbers have no context. “Eighty” and “eighteen” carry the same information load, so a single error changes the whole outcome. Mobile numbers are ten digits of pure, unrecoverable information.
Names, especially surnames
Given names are usually fine. Surnames are where you find out how good your setup is. A model will happily transcribe a rare surname as the nearest common English word it knows. Then your CRM creates a duplicate contact, the follow-up text goes to a name the customer does not recognise, and the whole thing feels cheap.
Suburbs and street names
Auckland alone has hundreds of suburb names that no international model has ever encountered in volume. If your system routes jobs by area, or quotes travel differently for out-of-zone work, a misheard suburb is a mispriced job.
Email addresses said out loud
Stop asking for these. Even a person struggles with a spoken email address on a mobile connection. Two callers in a row will say “dot co dot en zed” at different speeds; one will spell their surname with the New Zealand “zed” that some models read as the letter “said”, and you will collect a broken address that bounces silently. This is a design fault, not an accent fault, and I will come back to how to avoid it entirely.
Background noise, which we have a lot of
Our customer base is disproportionately mobile. Tradies calling from a site. Reps calling from the car. Café owners calling from the floor at 11 am. Noise degrades transcription accuracy on every accent, and it compounds an existing accent gap. A system that handles a quiet office call perfectly can drop badly at 100kph with the window cracked.
Turn-taking, which people mistake for mishearing
This is the sleeper problem. A meaningful share of “it kept cutting me off” complaints are not transcription issues at all. They are endpointing, the system’s decision about when the caller has finished speaking. Kiwis pause mid-sentence and finish statements on a rising tone. A system tuned for flatter, faster-finishing speech will decide you have stopped and start talking over you. The caller then repeats themselves, gets interrupted again, and hangs up convinced the AI could not understand their accent. It understood them fine. It just would not let them finish.
Fixing the AI Receptionist, Not the Kiwi Accent
Here is the part that matters. Every issue above has a specific fix, and none of them involves asking your customers to speak differently.
Pick a speech model that has heard us before
Transcription engines are not interchangeable. Some handle non-American English noticeably better than others, and the gap is measurable. Before you commit to a platform, get the same twenty test recordings run through the options and count the errors yourself. This is a twenty-minute exercise that will save you a month. Anyone who cannot show you accuracy on your own audio is asking you to take it on faith.
Load your vocabulary in before day one
Nearly every serious voice platform supports custom vocabulary or keyword boosting, which tells the model that certain words are far more likely in your context than the training data suggests. This is the single highest-value hour of setup work for a New Zealand business, and it gets skipped constantly.
What to load: every suburb and town you service, every product and service name you use, your team’s first names, common local surnames, your street address, and any te reo terms your callers use. If you are a dental practice, load the procedure names. If you are a plumber, load “califont”, “spouting”, “cylinder”, “gib”. Local trade language is as foreign to an American model as a place name is.
Confirm everything that costs money to get wrong
This is the discipline that makes an imperfect system reliable. Anything the AI is going to write into your system gets read back to the caller before it is committed.
Phone numbers get read back in pairs. Names get spelt back. Appointment times get confirmed with the day of the week attached, because “Tuesday the fourteenth at ten” gives the caller two chances to catch an error, where “1000 on the fourteenth” gives them none. Addresses get repeated with the suburb.
A system that is 92% accurate on raw transcription but confirms every critical field behaves better in practice than a 99% accurate system that confirms nothing. The confirmation loop is worth more than the model upgrade.
Take the mobile from caller ID and text the rest
The caller’s number arrives with the call. Use it. The AI says, “I have your number as 021 double four seven; is that the best one for a confirmation text?” That is one short confirmation instead of ten digits of dictation.
Then never ask for an email out loud. Text the caller a link and let them type it. You will get a clean address every time, you have opened an SMS thread you can follow up in, and you have removed the most error-prone thirty seconds of the call.
Tune the interruption settings for how we actually talk
Lengthen the silence threshold before the system decides the caller has finished. Give it a slightly longer patience window than the default. Then listen to ten real calls and adjust again. Most platforms let you set this per agent, and most installers leave it at the factory setting.
Give it a New Zealand voice, and let it own being AI
This is about trust rather than accuracy, but it changes call outcomes. A New Zealand, or at minimum Australian, voice keeps callers on the line longer than an American one, because an obviously offshore voice reads as an offshore call centre and people start looking for the exit. Name it, and have it disclose that it is an AI assistant when asked. Trying to pass it off as human is the fastest way to lose a customer who works it out at minute three.
Build the escape hatch
Two failed attempts at the same field, and the system stops trying. It takes a message, confirms the callback number, texts a link, and flags the call for a human. A system that gracefully hands off after two misses is not a failure. A system that asks a fourth time is.
How to Test an AI Receptionist on New Zealand Callers
Do not test this yourself in a quiet office with a good headset. You already know what the system is meant to hear, and you will unconsciously speak clearly for it. That test always passes, and it means nothing.
Run twenty calls before go-live. Twenty is enough to find the patterns and small enough to do in an afternoon.
Who calls: get five different people. Mix ages, mix strength of accent, and include at least one person with a strong regional accent, a Southland burr if you can find one, and one person who speaks te reo. Your own staff and a few friendly customers will do.
What they call about: the four or five things your customers actually ring about. Not edge cases. The bread and butter.
Conditions to include deliberately:
- One call from a moving vehicle with road noise
- One from a job site with a compressor or radio going
- One from a café at lunchtime
- One from a poor reception area where the line drops in and out
- One caller who talks over the AI on purpose
- One caller who gives a Māori suburb and surname
- One caller who says a time as “half ten” or “quarter to nine”
- One caller who changes their mind mid-booking
What you score: for each call, three things. Did it capture the details correctly? Did the caller feel heard? Would you have been comfortable if that had been a real customer? The third question is the one that matters, and it is the one people skip.
Then read the transcripts. Every platform worth using stores the transcript and the recording. You are looking for repeated substitutions, the same word going wrong the same way, because those are one-line fixes in your custom vocabulary. Ten minutes of transcript review usually generates a list of fifteen words to load, and that single change lifts accuracy more than anything else you can do in week one.
Keep reviewing weekly for the first month, then monthly. New suburbs, new products, and new staff names all need adding as the business changes. This is fifteen minutes a month, and it is the difference between a system that gets better and one that quietly degrades.
What Happens When the Phone Actually Gets Answered
The reason any of this is worth the effort is that the missed call is one of the most expensive things in a small business, and almost nobody measures it.
Dr Claire runs a dental practice with two receptionists at the desk. When we looked at the call data, 47% of inbound calls were going unanswered. Not because the team was slack, but because they were with patients, on another line, or at lunch. After putting an AI receptionist on the phones to handle overflow and after-hours, missed calls went to zero and booked appointments rose 44%. The receptionists did not lose their jobs. They stopped being interrupted every four minutes.
Justin Touyz saw a 27% revenue lift in the first month after deployment. Donna Loeffler doubled her sales in the month it went live. Different businesses, same mechanic underneath: the calls were always coming; they just were not being answered.
There is a well-known finding from lead response research that roughly 78% of deals go to the business that responds first. Every unanswered call is a customer who is now dialling the next name on the list, and they will not ring you back. That is what an unanswered phone call costs, and it costs you every single day whether you are looking at it or not.
The accent question matters because it decides whether you get to have that outcome. A voice system that mishears one caller in four does not just annoy people; it destroys the trust that makes the whole thing work. Get it configured properly for New Zealand callers and the phone stops being a leak.
The bigger point is that answering the phone is one job. If your business is like most I work with, there are another fifty recurring tasks like it, each one of them a small drain that only you or one key person can handle. The phone is a good place to start because the result is immediate and easy to measure. It is not where it stops. What actually changes a founder-led business is when the AI brain that holds your context and the AI workforce that does the doing cover enough ground that you stop being the bottleneck.
If you want a wider view of that, the pieces on AI agents for small business and admin overload cover the same logic applied beyond the phone.
The Short Version
An AI receptionist mishears Kiwi callers for reasons that are well understood and entirely fixable. Our vowels sit in a place the training data does not cover well, our place names are absent from it altogether, and default turn-taking settings are tuned for a different rhythm of speech.
The fixes are unglamorous. Choose a model that performs on your own audio. Load your suburbs, names and trade language as custom vocabulary before the first live call. Confirm every field that costs money to get wrong. Take the mobile from caller ID and text anything complicated. Lengthen the interruption window. Give it a local voice and a clean handoff to a human after two failed attempts.
Do those six things and the accent stops being the story. The story becomes a phone that gets answered at 7pm on a Saturday by an AI receptionist with a Kiwi accent, which is when half your customers were trying to reach you anyway.
If you want to work out whether this is the right first move for your business, or whether something further up the chain is the bigger leak, book a 30-minute discovery call. No pitch deck. We will look at where your calls are going and what it is costing you, and if the phone is not your biggest problem, I will tell you that. If you would rather read first, after-hours call answering and AI appointment booking both go deeper on the phone side specifically.
Frequently Asked Questions
Why does my AI receptionist mishear New Zealand callers?
Usually because the speech model was trained mostly on North American and British recordings, and New Zealand English has raised short front vowels that shift how numbers and common words sound. On top of that, local suburb names, te reo place names and trade terms are absent from the training data. Both problems are fixed at setup by choosing the right model and loading custom vocabulary.
Can an AI receptionist understand a Kiwi accent well enough to book appointments?
Yes, when it is set up for it. The reliability comes less from raw transcription accuracy and more from design: reading numbers back, spelling names back, confirming the day and time together, and taking the mobile from caller ID rather than dictation. A well-configured system books appointments accurately all day. A default configuration will not.
Does the AI receptionist need a New Zealand accent itself?
For accuracy, no. The voice speaking and the model listening are separate. For results, it matters a lot. Callers stay on the line longer with a local or Australian voice, because an obviously American voice reads as an offshore call centre and people start looking for a way out. Use a New Zealand voice if the platform offers a good one.
How does an AI receptionist handle Māori place names and surnames?
Only if you load them in. Nearly every voice platform supports custom vocabulary or keyword boosting, which tells the model that specific words are far more likely than its training data suggests. Load every suburb and town you service, common local surnames, and any te reo terms your customers use. Skipping this step is the single most common setup failure I see.
What should happen when the system genuinely cannot understand a caller?
It should stop trying after two attempts at the same detail. A good setup takes a message, confirms the callback number it already has from caller ID, sends a text with a link, and flags the call for a human to pick up. Asking a caller to repeat themselves a fourth time is worse than admitting the limit and handing off cleanly.
How long does it take to tune a voice system for New Zealand callers?
The initial vocabulary load and confirmation logic is a few hours of work. Testing with twenty real calls takes an afternoon. After that, budget fifteen minutes a month reviewing transcripts for repeated errors and adding new suburbs, products or staff names. Most of the accuracy gain arrives in the first week, then it improves slowly with review.
Will my customers know they are talking to an AI?
Some will, and that is fine. Give the assistant a name and have it confirm it is an AI assistant when asked directly. Trying to pass it off as a person is the fastest way to lose someone who works it out mid-call. In practice, customers care far more about being answered quickly than about who answered.
About Octavius
Titus Mulquiney is the founder of Octavius AI, where he builds AI brains and AI workforces for founder-led businesses stuck running everything out of their own head. Twenty years in marketing, ex-Sony product manager, ex-GM Zeal NZ. Based in Auckland, working with operators across NZ, Australia, and the US. Connect on LinkedIn.