Which voice should your agent use?

Choose the voice on a real call

Weeks later, your new AI voice agent calls a customer in the voice the meeting picked. She is at a bus stop with traffic going past, the voice starts reading her delivery time, and she hangs up before it gets to her address.

If you are about to choose an AI voice for calls to your customers, choose it the way they will hear it, on a phone, over a whole call, and at the moments where your calls get hard. Decide on what you cannot change later, the character of the voice, and leave pronunciation and pace for afterwards. Then put your two best voices on the same real calls and keep the one more customers stay with until they agree to what the call asks, such as a delivery time.

Why does the demo voice mislead?

A demo page is built to show each voice at its best. Every text-to-speech voice on it reads a demo sentence the vendor chose, about six seconds long and as clean as a studio recording, and the meeting hears it in a quiet room.

A customer hears the voice for your AI agent through a phone line, which carries less of the voice than a laptop or headphones do, and may be standing on the street or in a shop while it talks. The call lasts two or three minutes, and the voice reads the customer's own details instead of a polished line: their name, a time window, an address and an order number. When the customer says "Sorry?", the voice has to say it again.

Some voices that win in the meeting lose on that line. A bright voice with a lot of breath in it can sound wonderful on a laptop and turn thin on a phone, more so outdoors. An announcer voice, the polished kind you know from radio ads and audiobooks, can stay perfectly clear on the phone and still sound like a recording, and people hang up on recordings.

The voice is only one part of the gap between a demo and a live call; our post on agents that pass the demo and struggle on live calls covers the rest.

What should the voice sound like?

Language comes first. The voice should be a native speaker of your customers' language and of the regional variety they speak. Portuguese from Brazil and Portuguese from Portugal are heard as different at once, so a voice that suits customers in São Paulo is the wrong voice for customers in Lisbon, however good it sounds.

Then think of the rep your customers are happiest to stay on the phone with, and look for an everyday voice like theirs, the sound of someone on your own team talking rather than an announcer voice. Whichever voice you choose, set it to an even pace that slows a little for times and numbers, so the customer can write them down.

The voice also has to fit the persona, the name and manner your agent introduces itself with. An agent that calls itself Mia needs a woman's voice. Sounding natural is not the same as pretending to be a person, and where your market requires it, the agent says in its greeting that it is an AI, whatever voice it speaks with.

Which moments should you test?

Consider [store], an online furniture store. The store, the people, the panel, the scores and the calls in this post are an example, not real data. The day before each delivery, [store]'s AI agent calls the customer to confirm the time window and the address. Lena, who runs customer service at [store], has a shortlist of three voices from the vendor's demo page, all women's voices because the agent introduces itself as Mia. In the meeting, everyone liked Voice C best, a bright voice with a lot of breath in it.

Start, as Lena does, with the call your agent makes most often. She writes [store]'s delivery call out a little longer than usual, so that it runs about three minutes with every hard moment in it, and plays this test call through each voice on her own mobile phone, once in her office and once outside on the street.

The test call opens with "Hello, this is Mia, the AI assistant at [store]." Mia checks that she is speaking to Mrs. Szczepańska, a hard surname from [store]'s customer list, and tells her that her sofa arrives "tomorrow, Thursday, between 9:00 and 12:00", at 14 Mill Lane. Mia reads the order number in two groups, 4471 2093. The customer says "Sorry?", and Mia reads it again, more slowly. The customer then asks for the afternoon, which is full tomorrow, so Mia says she is sorry and offers Monday between 13:00 and 16:00. The customer takes it, Mia reads back the day, the time window and the address, and they say goodbye.

A demo sentence lasts about six seconds, about as long as the greeting. The meeting judged the voices on that first moment and heard none of the other seven.

How do you score the shortlist?

Lena puts together a listening panel of five people: two reps who talk to [store]'s customers every day, herself, a delivery coordinator who knows what customers complain about, and one person from outside the project. They listen blind, to Voice A, Voice B and Voice C, with no vendor names and no prices. Each of them hears the whole test call on a phone and scores every voice from 1 to 5 on five questions.

All three voices read the test call with the same settings, the same pace and the same grouping of numbers, so any difference the panel hears comes from the voice itself. In the table "How the panel scored three voices", each cell is the five listeners' scores added up, so a cell runs from 5 to 25 and a total from 25 to 125.

Voice C, the favorite from the meeting, came last on the phone with 75 of 125. It was the lowest of the three on being easy to follow on the phone, 12 of 25, and on clear numbers and times, 13. The brightness that sounded so clear on the laptop thinned out on the phone line.

Voice A, the announcer voice on the shortlist, was the clearest of the three on numbers and times, with 21 of 25, and the lowest at minute three, with 11. It also scored only 14 on sounding like someone from here, and two listeners wrote "sounds like an ad" next to it.

Voice B scored 103 of 125 and came first on four of the five questions. Voice A beat it only on numbers and times, 21 of 25 against 20.

The row for names and brand came out nearly even, 18, 19 and 18 of 25, because all three voices got Mrs. Szczepańska's surname wrong in the same way.

Lena takes the two highest totals, Voice B and Voice A, to a live test. Two listeners heard an ad in Voice A, and only real calls can show whether customers hang up on it.

What can you fix later?

Choose on the character of the voice, because the settings can be changed afterwards, whichever voice you pick. Pronunciation rules set how the voice says names, places, your brand and your product terms, and the pace, the pauses and the way numbers and order numbers are grouped are settings too.

That is why the names row in Lena's table should not decide anything. The surname went wrong in pronunciation, and one rule puts it right in any of the three voices. Writing such rules is the subject of our post on pronunciation dictionaries for names, brands and products.

Character comes with the voice. How it sounds through a phone line, whether it sounds like an announcer and whether it is still easy to listen to at minute three stay the same whatever you set. Slowing Voice C down will not make it fuller on a phone, and no setting stops Voice A from sounding like an ad.

Which voice do customers stay with?

For two weeks, every customer on [store]'s delivery list gets Voice A or Voice B at random, with the same script and at the same hours, so the voice is the only thing that changes. By the end of the two weeks each voice has 400 answered calls. Lena does not compare earlier, because a few dozen calls each would be too few.

She sorts every answered call into one of three outcomes that add up to 400. A customer who hangs up within the first ten seconds is an early hang-up, and one who agrees to a time window, the planned one or a new one, has confirmed a time. Any other ending, for example when the customer asks for a person, the line drops or no time is agreed, goes in the column "Other". Lena also counts the customers who asked to repeat at least once. A customer can do that and still confirm a time, so the column "Asked to repeat" is not part of the 400.

With Voice B, 338 of 400 customers confirmed a time, against 296 with Voice A, so 42 more customers agreed to a delivery time. Voice B also had fewer early hang-ups, 44 against 72, and 18 of its calls ended another way, against 32 for Voice A.

Every customer who asked to repeat had stayed past the first ten seconds, so the fair comparison is with those customers: 36 of 328 for Voice A against 41 of 356 for Voice B, about the same share. They followed the announcer voice about as well, and still more customers hung up on Voice A in the first ten seconds.

Here the customers agreed with the panel, and Lena keeps Voice B for all of [store]'s delivery calls.

Where does DRING come in?

In DRING's speech core, each brand picks a curated voice from the library for each language, with a persona for pace, formality and tone, accent correction and a pronunciation dictionary for names, places and product terms. You approve the wording, the persona and the voice before anything is dialed.

Which opening line, tone or voice works is settled on live outbound calls, not in a meeting. In DRING's outbound campaigns, two voices can run side by side on the same campaign to compare results. The voice samples on our home page are DRING workflow demos; play them on your phone rather than on your laptop.