Choosing a voice for your AI agent: pace, warmth and language
The voice is the first thing a caller notices and the last thing most people choose. They spend a week on the instructions, ten minutes on the voice, and then wonder why callers hang up in the greeting. On the phone the voice is the uniform: it tells the caller what kind of place they have reached before a single fact is exchanged. Here is how to choose one on purpose, and a five-minute test that catches most mistakes.
Start from the caller, not the catalogue
Voice catalogues are long and every sample sounds fine in a headset. Ignore them for a moment and describe your caller instead. A patient calling a clinic at seven in the morning wants calm and clear. A tradesman calling a builders' merchant wants quick and plain. A guest calling a hotel wants warm and unhurried. A buyer calling an estate agent wants confident.
Write that one sentence down. Then describe your best human receptionist in the same way. Most businesses find that the voice they want is the one that sounds like the person they would hire, not the one that sounds most impressive.
Pace
Phone callers cannot rewind. A voice that reads a phone number at the speed of a podcast host makes people say "sorry, what?", and a voice that reads everything at the speed of a bedtime story makes them tap the desk. The right pace is a little slower than conversation, and slower still on numbers, addresses and times.
Every voice has a natural speed, and the instructions can nudge it. Ask for short sentences and the agent will sound quicker without speaking faster. Ask it to read numbers in groups of two and to pause after a date, and the slow parts land where they are needed. Test with a long address, a phone number and a time, not with a greeting.
Warmth, character, gender
Warmth is the difference between a voice that seems to smile and one that reads. Most businesses want some; very few want a lot. A very bright, upbeat voice tires callers within a minute, and it grates on anyone ringing about a problem. A flat voice sounds like a machine even when the words are kind.
Gender and age have no rule, and the honest advice is to avoid deciding on a stereotype. Choose what your callers expect from your business and what your staff sound like, so the handover to a human does not feel like a jump. Pick one and keep it: a business that changes its agent's voice every month sounds like a business that cannot decide who it is.
Listen to each candidate say something that is not a greeting: a "no", a "let me check that", and an "I'm sorry, that time is taken". Voices that only sound good delivering good news will let you down on the calls that matter.
Language and accent
A voice must be made for the language it will speak. A voice built for English and asked to read Danish, Swedish or Norwegian sounds like a well-meaning tourist, and callers hear it in the first sentence. The same is true the other way round: a Danish voice reading an English street name. Our posts on Danish and on Swedish and Norwegian list the words that give a voice away.
NordTell offers voices from ElevenLabs and Cartesia, and the provider is printed next to each one. Both have voices made for many languages; the way to choose between them is to listen, in your language, to your ten most common place names and a phone number. If callers may switch languages mid-call, choose a voice that handles both languages you expect, and read what happens when a caller switches language mid-call.
The engine matters too. The Pipeline engine has the widest choice of voices; speech-to-speech has fewer, because the voice is part of the model. If a particular voice is important to you, check which engines can use it before you fall in love with it.
The five-minute browser test
Before the voice meets a caller, make it say these lines to you from the browser, and listen on a phone speaker rather than headphones, because that is what callers use.
- The greeting, including the line that says it is an AI.
- A phone number read back in pairs.
- A full address with a floor and a postcode.
- A date and a time, including a "half past".
- A polite refusal: "I'm sorry, that time is taken. Would half past ten work?"
- A transfer line: "Let me check if she's available, one moment."
- Three place names your callers say most.
Then ask a colleague to listen without telling them which voice you prefer. If you both wince at the same line, change the voice; if you wince at different lines, change the instructions. Our guide to testing a voice agent before launch covers the rest.
Where NordTell fits
In NordTell the voice is a setting on the agent, next to the language, the engine and the instructions, and you can change it and call the agent from the browser as often as you like on a free account. Voices come from ElevenLabs and Cartesia, with the provider shown next to each. The test above takes five minutes and needs nothing but a microphone.
How agents are set up · Pipeline, speech-to-speech or hybrid?