AI receptionist voice and accent: what keeps callers on

Pick a clear voice with an accent from your callers' own country. Give it a job title instead of a human name. Make its first sentence say who they've reached, that it's an AI assistant, and what happens next. Then make it fast. The accent matters less than you'd think. Callers hang up when the first five seconds give them nothing, when the line goes quiet, or when the voice pretends to be a person.

I'm a plumber by trade, and I build AI phone receptionists for plumbing, HVAC and electrical firms in the US and the UK. My first client, a plumbing company, goes live this month, with a script written from real call transcripts. What follows is the research I read and the choices I'd make. It isn't caller data, because I don't have any to publish yet.

Callers hang up on the first five seconds, not the accent

In a field experiment published in Marketing Science in 2019, a lender in Asia had either an AI voice bot or a person make sales calls to 6,255 customers about renewing a loan. When the bot opened with "Dear customer, I am the AI voice chatbot of the company XYZ", 56.3% of the people who answered hung up within five seconds. Sales fell to 4.8%, against 23.7% when the same bot never said it was AI. When the same bot owned up to being AI after the pitch instead, 4.5% hung up.

Before you use that to hide the AI, look at those calls. The bot rang them, selling. Nobody needed anything. Your caller rang you because something is leaking, tripping or not heating. What carries over is the mistake: a first sentence all about the machine gives the caller a reason to leave and none to stay.

The other killer is silence. Across ten languages, researchers timing answers to yes-or-no questions in real conversations found the average gap before a reply was about 200 milliseconds, and the typical English reply came with no gap at all. A one-second pause on the phone is long. At two seconds the caller says "hello?" and wonders if the line dropped.

Here's what a weak first five seconds can cost, as a worked example, not a measurement. Say 60 calls a month reach the AI: after hours, plus the ones you can't grab on a job. Say a clumsy greeting loses one caller in ten before they've said what's wrong. That's six callers. If one in four would have been a job worth $450 (about £350), that's one or two jobs a month, about $8,100 a year. Your numbers will differ, and the calculator runs them.

Pick an accent from your callers' country, and use their words

The biggest UK study of accent attitudes I could find is Accent Bias in Britain, from Queen Mary University of London and the University of York in 2020. They had 827 people rate 38 accent labels. Received Pronunciation still came top for prestige and Birmingham bottom, as in surveys 15 and 50 years earlier. Two findings matter more for your phone. "Own accent" made the top ten. And when 1,062 people heard real voices giving mock job-interview answers, the differences between accents were much smaller than the labels suggested. Stronger answers lifted every accent's score. What the voice says counts for more than how it sounds.

So here's my rule. Same country as your callers, clear on a bad line, and a light local accent if the platform has a good one. A neutral American voice for a US firm. A British voice for a UK one. If you're in Leeds and there's a decent Yorkshire voice, try it against a neutral one on a few customers. What I'd never do is answer a UK number with an American voice, or the other way round. It sounds like a call centre somewhere else, and I don't want a caller wondering who's really going to ring them back.

Test the local voice before you trust it. A 2025 study of ElevenLabs and Speechify found their voices performed unevenly across five regional English accents. A synthetic local accent can come out as a caricature, and a caller from that town will hear it first.

The words matter more than the vowels. A voice that asks a Manchester caller for their zip code has lost them, whatever its accent. Boiler or furnace, consumer unit or breaker panel, call-out charge or service call fee: the script has to speak the caller's trade English. I went through the UK list in the UK plumbers post.

The accent that matters most is the one it hears

You choose the voice it speaks with. You don't choose the voices it has to understand. That's where accents actually cost jobs.

In 2020, Stanford researchers ran interviews through five big speech-to-text systems, from Amazon, Apple, Google, IBM and Microsoft. They got about one word in five wrong for white speakers and one in three for Black speakers, and the error rate climbed the more dialect features a speaker used. Those systems have improved since. The lesson hasn't changed. Speech recognition learned from some voices more than others, and your callers don't get to pick theirs.

This is where I tell you what AI can't do. It will mishear a street name. It will get a letter in a postcode wrong. It won't know that Mrs Patel has rung you for twenty years and always says "the flat above the shop". So the fix isn't a better voice. It's the script and the test:

• Read back the important bits. Zip code or postcode, street, callback number, digit by digit. The caller corrects it on the spot.
• Ask one thing at a time. A long question answered in a strong accent, over a pump running, is where it slips.
• Test with real voices. Your apprentice, your mum, the customer who always rings from site. Not you in a quiet office.
• Get it to a person fast. When it mishears, the human calling back in minutes catches it.

A fuller test list is in the voice-agent testing checklist.