Cloning an employee's voice with AI: it's legal, but only if you do three things first

Google now lets you clone a voice from a 30-second audio sample ¹. The feature arrived with the Gemini 3.8 Flash TTS models, released on 23 September 2026 ¹. Google itself states the condition: it must be your voice, or a voice you have the right to use ¹.

That sentence sums up everything you need to know. Cloning the voice of an employee or a spokesperson is not forbidden. It becomes a problem when you do it without a written agreement with the person, or when listeners don't know they are hearing a synthetic voice.

Why this is now your problem, not just for podcasters

Until recently, cloning a voice took a sound engineer, days of work and a budget. Today it takes a call to an online service.

Simon Willison, a developer who tries these tools as soon as they come out, generated 1 minute and 18 seconds of dialogue in about 20 seconds, spending 2.74 cents ¹. At costs like that, sooner or later someone in your company will try it. The marketing manager who wants an advert without a recording studio, or whoever looks after the switchboard and wants a warmer hold message.

It's better that you set the rules first, before someone uploads a colleague's WhatsApp voice note and hits "generate".

A person's voice is not a file like any other

Under the GDPR, a recording that lets you recognise someone is personal data. The model you build from it rests on that data, so you need a legal basis to use it and you must tell the person what you are doing with it.

If the voice were used to identify the person, for example to unlock access, it would become biometric data with even stricter protections. For a switchboard or an advert that is usually not the case, but it doesn't give you a free hand.

There is also a level that is often forgotten: a voice is a recognisable trait of the person, like the face. Italian case law tends to treat it much like their image, so using it for commercial purposes without permission is a risk even outside the GDPR.

With an employee, a verbal yes is worth nothing

This is the most common trap. Your employee says "of course, no problem" by the coffee machine. You go ahead. You're in a bad position.

Between employer and employee, consent is considered weak, because someone who depends on you rarely feels free to say no. That is why the agreement must be separate from the employment contract, genuinely optional and in writing. At the very least, put these things in it:

• which uses the cloned voice covers (switchboard, radio advert, social media video) and which it doesn't;
• who approves the texts the voice will say;
• whether there is payment, and how much;
• what happens if the person withdraws consent or leaves the company;
• where the voice model is kept and who can use it.

The point about leaving the company is the one that does the most damage. Imagine a customer who calls and hears the voice of someone who now works for a competitor. State from the start that, when the relationship ends, the model is deleted and the recordings are removed from your channels within a set time.

What the AI Act asks of anyone who makes a synthetic voice speak

The AI Act has a transparency chapter that concerns you for two reasons.

The first is the switchboard. If an artificial intelligence system talks to people, they must know they are interacting with a machine, unless it is obvious. A cloned voice that sounds exactly like the lady in accounts is the opposite of obvious.

The fix costs one sentence: "You are talking to an automated assistant". Put it at the start of the call, not the end.

The second is advertising. Generated audio that imitates a real person falls within the content the AI Act asks you to declare as artificial. In an advert a short notice is enough, in the spoken text or in the caption. It must not be hidden.

The timetable for applying the AI Act has been much debated. Don't build a strategy on it of the "I'll wait until it becomes mandatory" kind: declaring a synthetic voice costs you nothing, and saves you the embarrassment when a customer notices on their own.

The spokesperson is a different case, and a more delicate one

With an external spokesperson, whether a well-known local face or a satisfied customer, the problem shifts to the contract and to misleading advertising.

If you clone their voice, you can make them say anything. For exactly that reason the contract must say which sentences the voice may speak, who approves them, for how long and on which channels. A licence to use a voice should be written with the same care as a licence for a photo.

And never make them say things they don't think. A spokesperson who "recommends" a product they have never tried, in their own real voice, is misleading advertising even if you have a signed contract.

What not to do, plainly

Don't clone the voice of a former employee, even if you still have the old switchboard recordings. You held those recordings for another purpose.

Don't ask for a voice that "sounds like" an actor or a famous character. If people recognise it, the problem is the same as having cloned it.

Don't have the consent signed together with the hiring contract or the renewal. It looks like blackmail, and a judge would read it that way.

What to do now

First ask yourself whether you really need a cloned voice. Gemini offers a library with over 2,000 ready-made voices ¹. For a switchboard or a tutorial, a library voice works perfectly well and removes every consent problem. In most cases that is what I recommend.

If instead the recognisable voice has real value, for example because it is the owner who has been speaking to customers for years, proceed like this:

1. write a separate agreement, with uses, payment and exit;
2. update the privacy notice and the record of processing activities;
3. put the "synthetic voice" declaration at the start of every call and in every advert;
4. decide who in the company can generate audio with that voice, and keep track of what is generated.

If you want a second opinion before signing, we at AIR are happy to look at it. But you can apply the basic rule yourself from tomorrow: a cloned voice still belongs to the person who lent it, and you are only renting it.