Our first outbound campaign did not work until we stopped using a synthetic voice and started playing recordings made by a professional voice actor in a studio. Then conversion appeared. That was the first time we saw that in a cold call, people hear the voice before they hear the offer. Ten years and tens of millions of conversations later, we no longer think recordings always win. We do think the voice is a business variable, and it deserves the same scrutiny as the script, the schedule and the database.

This is an opinion piece, but it rests on practice rather than taste. Below we explain what happened in that first campaign, why a cold call is the hardest possible test for a voice, what a studio voice really involves, what has changed since speech synthesis became good, and how we decide today which voice a given process should use.

The campaign nobody ordered#

In the years after our founding in 2012 we were known, where we were known at all, for understanding speech rather than merely recognising it. Inbound conversations were the natural home for that: a person calls, says what they need in their own words, and the system works out the intent. Outbound was different. Nobody asked us to try it. We designed our first outbound campaign ourselves and went to test it with a single call centre.

The campaign asked people whether they were interested in a new mobile-internet bundle. Two decisions shaped it. First, before dialling anyone we checked the phone numbers, so that attempts went only to numbers worth calling. Second, after the first runs with synthesis, we replaced the synthetic voice with recordings by a professional voice actor. That second change is the one this article is about. With synthesis, the calls produced little. With the studio voice, we suddenly had conversion, and not a trivial amount of it.

We did not run a formal experiment with control groups; we were a small company learning by doing. But the difference was large enough to change how we worked. Our first large campaign followed shortly afterwards: confirming people’s interest in a loan. It ran on recorded voices. So did most of what came next.

Why a cold call is the hardest test for a voice#

Clients often assume that outbound is the easier half of telephone contact. In our experience it is the opposite, at least if the goal is effectiveness rather than activity.

Clients often assume outbound is simpler than inbound. Someone who calls you will cope even with an IVR. Someone you call may not even want to pick up, let alone keep talking. – Dawid Wójcicki, Founder & CEO of Vee

A person who calls a helpline has a reason to persist. They will press keys, repeat themselves and forgive a stilted prompt because they want something. A person you call has no such motive. They were doing something else. They did not expect you. In the first seconds they decide whether this call is worth their attention, and the sound of the voice is a large part of that decision, often larger than the words.

That is why we measure cold contact on a ladder rather than a single number. Reach comes first: without reaching the customer, there is nothing else. Then continuation, the inverse of the share of people who hang up as soon as they hear the company name. Then obtaining the key information, such as interest, a confirmation or a promise to pay. Then the real business effect, which happens in the client’s systems. The voice acts mostly on the second rung, and the second rung decides whether the third and fourth are reachable at all.

  1. Reach. Did we get through to the right person?
  2. Continuation. Did they stay on the line after the opening sentence?
  3. Key information. Did we obtain the interest, confirmation or promise the process exists for?
  4. Business effect. Did a payment, contract or attendance follow in the client’s systems?

In 2016, synthesis lost on rung two. The listener heard a machine reading text, and the most common response to a machine reading text is to hang up. A studio voice did not hide what the call was; it simply did not give the listener an immediate reason to leave.

A studio voice is not a disguise#

One misunderstanding needs clearing up before going further. Choosing a natural-sounding voice has never been, for us, a way of pretending to be human. In May 2016, when we debuted on NewConnect, we told Puls Biznesu that we would not hide from customers that they were talking to a robot. That has stayed our rule: Vee always introduces itself as AI.

Since 2 August 2026 this is also an explicit legal requirement. Article 50 of the EU AI Act says people must be informed that they are interacting with an AI system, and the European Commission’s FAQ on Article 50 adds that the notice must come from the start of the first interaction and be clear and distinguishable. The Commission published guidelines on these transparency obligations on 20 July 2026. Nothing about a good voice conflicts with that. A listener can know perfectly well that they are speaking to an AI and still find the conversation pleasant, clear and worth finishing. The voice decides whether they stay; the disclosure decides whether they can trust what they stayed for.

What a studio voice actually involves#

“Recorded voice” sounds simple. In practice it is a craft, which today sits in our voice studio, Vee Lector/UX. A professional voice actor records every phrase the dialogue may need, in several intonations where the situation calls for it. A question sounds different from a confirmation; an apology sounds different from a reminder. The recordings are then joined at run time, a technique called concatenation, so that a sentence containing a date, an amount or a product name still sounds like one sentence spoken by one person.

Most of the work is in decisions the listener never notices. Which word in a sentence carries the stress. How long the pause is before the key question. Whether a question is open or closed. What emotion fits a particular moment. Which vocabulary suits the people being called, because a phrase that sounds natural to a 30-year-old city customer may sound odd to a retired person in a small town. Tuning these choices is ordinary work for our dialogue designers and voice studio, process by process.

The trade-offs are real#

  • Every new phrase needs a studio session. Changing a sentence is not a text edit; it is a recording, an acoustic check and a release. That slows down changes.
  • Variable content is harder. Names, amounts and dates must be joined so the seams are inaudible. Bad concatenation is worse than decent synthesis.
  • Coverage is finite. The dialogue can only say what has been recorded. That is a limitation for open-ended conversations.
  • The same finiteness is an asset in regulated work. A closed set of recordings is a countable, auditable catalogue of everything the system can say.

The last point matters more than it first appears. In a bank’s early-collections process or an insurer’s renewal campaign, the client’s compliance team wants to know exactly what will be said to customers. With a closed catalogue of recorded utterances, they can read and hear every one of them, approve them, and see each change versioned. The voice choice and the governance choice turn out to be connected.

What changed when synthesis got good#

If this article had been written in 2016, its conclusion would have been short: use a studio voice. It is not 2016. Speech synthesis has improved beyond recognition, and the best synthetic voices today are expressive, fast and available in many languages. We use synthesis ourselves, and our generative Brilliance 4 variant pairs generative mechanisms with modern synthesis for low-sensitivity processes, though it is technically still being built.

So the question has changed from “which technology is better” to “which voice performs best in this process, for these people, under these rules”. Sometimes that is still a studio voice. Sometimes it is synthesis. Often it is a hybrid of the two. The answer depends on the process, and it can change during the life of the process.

The speed of change matters as much as the sound. A synthetic phrase can be rewritten and released the same day; a recorded one needs the voice actor back in the studio. In a process where wording changes every week, such as a campaign that follows a client’s shifting product offers, that difference can outweigh a small advantage in how the recording sounds. In a process where the wording is approved once by a compliance team and then left alone for months, it hardly matters. We weigh the expected rate of change alongside the expected effect on listeners, and we tell the client which of the two is driving our recommendation.

How the voice decision tends to look by type of process
ProcessWhat the voice must achieveWhere we usually start
Cold sales contact and lead pre-selectionKeep a person who did not expect the call on the line past the first sentenceStudio recordings; cold contact often converts better with them than with synthesis
Early collections in banking and consumer financeSound calm and credible, say exactly what compliance approvedClosed, versioned catalogue of recorded utterances; synthesis only where approved
Appointment confirmation and reschedulingHandle many dates, times and locations clearlyHybrid: recorded frame, carefully joined or synthesised variable content
Surveys and feedbackStart every call the same way, no matter how many refusals came beforeEither works; consistency matters more than the technology
Low-sensitivity inbound questionsAnswer varied questions quicklySynthesis is often the practical choice

Starting points from our operating practice, not fixed rules. Each process is tested and adjusted after launch.

How we decide: test it, on the right rung#

We treat voice the way a good marketing team treats a subject line: as a hypothesis to test, not a preference to defend. Persona, voice and style are one of several parameters of a contact strategy, alongside numbers, timing, content and attempts. Each of them can move results. In our experience, conversion depends first on the voice, second on the content and only third on the infrastructure, which is a reason to take the first one seriously.

In practice, where the client agrees, we test voices against each other on a larger base, with everything else held equal. We compare on the rungs of the ladder that matter for the process. A voice that lifts continuation but lowers the share of promises to pay is not a better voice for collections. A voice that sounds pleasant in an internal demo but raises hang-ups after the company name is not a better voice at all.

Good voice tests are less glamorous than they sound. The design has to make the comparison fair, so that the result reflects the voice and nothing else, and the test has to run long enough for the slower rungs of the ladder to show up. And we agree with the client in advance which metric decides, so nobody is tempted to pick the winner afterwards by choosing the most flattering number.

Measurement is only half of it. Every day, in every project, people on our team listen to selected calls, especially unusual ones. Automatic metrics tell us that hang-ups rose on a Tuesday afternoon; listening tells us that a recorded phrase sounds clipped on certain phone networks, or that a pause before the key question feels like the line has dropped. Most of our ideas for improving a voice come from that listening, not from dashboards.

What tens of millions of conversations taught us about voice#

By 31 October 2026, Vee’s AI had conducted 38,325,799 connected conversations, cumulatively (Vee operating data). Every one of them was executed by Vee’s AI; Vee employs no call agents. That scale does not make us right about every voice decision, but it does mean our opinions have been tested many times against real listeners in banking, insurance, healthcare, telecoms and energy. A few lessons have held up.

38,325,799Connected AI conversationsConversations in which Vee’s AI connected with the person called or calling.Cumulative to 31 October 2026, Vee operating data.
  • The voice must match who is calling. We do not call on behalf of a bank; we call from the bank, as part of its own operation. The voice has to sound like that institution would sound on its best day.
  • The first sentence carries the most weight. Most hang-ups happen at the start. That is where the best recordings and the most careful pacing belong.
  • Consistency is an advantage. A human survey agent loses enthusiasm after thirty refusals. Our AI starts every call the same way, and a good voice keeps that consistency audible.
  • A voice cannot fix a bad base. If the numbers are stale or the people were never interested, no voice will rescue the campaign. We can adapt the process; we cannot jump over the database.
  • What worked last year may not work this year. Listeners change, networks change, competitors’ calls change what people expect. Voice decisions need revisiting.

Where we stand on the debate#

The market today talks a great deal about voices: how natural a synthetic voice sounds, how fast it responds, whether it can be interrupted mid-sentence. Those are real engineering achievements. They are also easy to demonstrate in a two-minute demo and hard to connect to any business outcome. We rarely see anyone publish what a given voice did to continuation, to promises to pay or to attendance in a specific process.

Our position is simple. Voice is a means, not a product. A studio voice won our first cold campaign because it kept people listening, and it still wins many cold contacts today. Synthesis wins in other places, and we use it there without sentiment. What we will not do is choose a voice because it sounds impressive to the buyer, or keep one because it worked five years ago. The choice belongs to the process and to the evidence.

What to ask a supplier about voice#

If you are evaluating automated telephone contact, the voice will be one of the first things you hear in a demo, and one of the easiest things to be impressed by. These questions move the conversation from impression to evidence:

  1. Which voice would you use for this process and these customers, and why?
  2. Can you run an A/B test of two voices on part of our base, and on which KPI will you judge the result?
  3. How do you measure hang-ups after the opening sentence, and will we see that number?
  4. If you use recordings, how quickly can a phrase be changed, and how is each change versioned and approved?
  5. If you use synthesis or generated speech, how do you ensure the system only says what our compliance team approved?
  6. How and when does the caller learn that they are speaking with AI?
  7. Who listens to real calls, how often, and what happens when they hear a problem?

A supplier who answers these with numbers from your own process, rather than adjectives about their technology, has understood that the voice works for the outcome and not the other way round. That has been our working assumption since that first campaign, and it is how we choose voices for the processes we run for clients today.

Follow The First Second

Occasional issues on the mechanics and economics of AI-powered customer contact. No noise, no vendor news.

Follow via RSS