How to choose a voice for your AI voice agent
Anyone setting up an Expedify voice agent who needs to pick the voice it speaks with. No technical setup required — this is a listening and judgement exercise, and it takes about ten minutes.
Your voice agent's voice is the first thing a caller hears, and most of what they'll remember. It carries more of your brand impression than the script does. This guide helps you choose deliberately rather than accepting the default.
Hear a premium voice first
Listen with the audio you'd expect a real caller to have — phone speaker, not studio headphones. A voice that sounds warm on a laptop can sound thin over a mobile connection, and that's the version your customers get.
What to listen for
Play the sample more than once and pay attention to something different each time.
Naturalness under pressure. Any voice sounds fine reading a simple sentence. The tell is what happens with the awkward material: a long number, an email address, an unusual surname, a price with decimals. That's where synthetic voices break character, and where your callers will notice.
Pace. Faster reads as efficient and can read as rushed. Slower reads as calm and can read as laboured. Match it to the call's purpose — an outbound sales call and a support callback want different tempos.
Warmth versus authority. A warmer voice suits nurture, retention, and support. A more measured, level voice suits qualification, collections, and anything where the caller needs to trust a number you're quoting.
Accent and locale. The voice should sound native to the people receiving the call. A mismatch is subtle but corrosive — callers can't always name what's off, and they trust the call less anyway.
Choosing for the job, not in the abstract
Different call types justify different voices, and you can run more than one agent.
| Call type | What the voice should do |
|---|---|
| Inbound support | Reassure. Warmer, unhurried, patient with interruptions. |
| Lead qualification | Sound competent and brisk. Neutral warmth, clear consonants. |
| Outbound sales | Sound like a person worth listening to. Energy without pressure. |
| Appointment reminders | Get out of the way. Neutral, clear, quick. |
| Collections or billing | Sound calm and factual. Never bright or apologetic. |
If you're only setting up one agent, pick for your highest-volume call type and revisit later.
Test before you commit
A sample reel is recorded under ideal conditions. Your script is not.
- Build the agent with your real opening line and one difficult passage — a number, an address, a name people mispronounce.
- Run a test call to your own phone. See Test your voice agent.
- Listen to the recording rather than the live call. You hear differently when you're not participating.
- Play it to someone who hasn't seen the script. Ask what they thought the company does, and whether they'd have stayed on the line.
That last question matters more than whether you personally like the voice.
Common mistakes
Choosing on the demo sentence. Demo lines are written to flatter the voice. Your script wasn't.
Optimising for impressiveness. The goal isn't a voice that makes people say "that sounded real". It's a voice nobody thinks about, because they were busy listening to what it said.
Ignoring the handoff. If your agent transfers to a human, the gap between the two voices is audible. A very polished AI voice followed by a rushed human on a headset can make the human sound like the downgrade.
Setting it and forgetting it. Voice choice is cheap to change and worth revisiting once you have real call recordings and outcomes to compare.
Next steps
- Build a voice agent — the full agent setup
- Voice channel overview — how voice fits with your other channels
- Transfer to a human — handing a call off cleanly