Voice cloning: from the Meloni case to audio advertising

Voice cloning makes it possible to generate new messages in a recognisable voice. For producers of commercials and jingles, this changes the project: they need to define what that voice may say and who controls subsequent versions.

The Giorgia Meloni case highlights the risk of associating someone with content they have never recorded. In advertising, the same issue affects production decisions: authorisation, delivery and responsibility for the final message.

The Meloni case: what was reported

On 5 October 2026, Associated Press reported that Giorgia Meloni had applied to the EUIPO for a voice trademark. The application included a four-second recording in which she repeated “Io sono Giorgia Meloni”. [1]

An application is not a granted registration. The report does not establish that protection against every form of voice cloning has been granted.

For audio production, the important issue is the connection between a voice and a person. Listeners may recognise who appears to be speaking and attribute the message to them. What is said, the product involved and the listening context all matter.

From a recording to a tool that generates further messages

When a voice artist delivers a recording, they deliver a defined performance: a particular script with a particular delivery. A voice model can be used to produce subsequent phrases.

This expands the possible uses. Before uploading material to a platform, distinguish permission to use the recording from the conditions for creating a model and generating further audio.

The agreement should also address subsequent uses: other clients, additional languages, sharing with suppliers and what happens to the model when the relationship ends. A general phrase such as “use of the voice” leaves these decisions open.

Recognisability and delivery are different things

Recognising the voice is only one part of listening. A commercial’s meaning also depends on emphasis, the pause before an offer and the way the closing line is delivered.

In a hypothetical example, “Today only, get the second product free” can emphasise either the deadline or the benefit. Both readings may be understandable, but they direct attention towards different information.

Judging only how closely the generated voice resembles the original is therefore an incomplete approach. Listen for whether the delivery supports the script and whether the offer remains clear.

A jingle also needs to work with the music. A different pause may move the brand name away from the intended musical accent. The useful check is on the final mix, with music and effects, in the format intended for distribution.

Voice direction remains a production decision even when a model generates the message.

Choosing a recorded voice, a synthetic voice or both

The choice should start with the work the voice needs to do.

For a campaign built around a specific performance, recording with the voice artist allows the producer to direct the delivery and discuss the script’s intentions during the session. For recurring messages with variable elements, an authorised model can be assessed alongside a process for reviewing each version.

A mixed approach is worth considering when the sonic signature must remain stable and only an informational segment changes. One option might be an approved recording of the brand name combined with an authorised synthetic voice for service updates.

This is a design option that needs a listening test. Rhythm, vocal character and sound treatment must make the segments coherent. If the change of voice distracts from the message, reconsider the approach.

The choice depends on the content’s purpose, with explicit control over how it is produced and updated.

A short variation can change the commercial promise

Imagine a thirty-second campaign featuring a promotional price. The next update involves changing just a few words.

Before regenerating the audio, check the complete sentence. Does the new price include the same conditions? Has the deadline changed? Does the shorter format still leave enough room to explain what the offer applies to?

Approval should cover both the updated script and the audio intended for distribution. Keeping them together helps establish which version was actually authorised.

The same applies to localisation. A voice that can speak Italian or Spanish does not by itself complete the adaptation: names, prices and promotional wording need to be pronounced correctly, and the script needs to work in the target language.

Carry out the language review before the final mix. Otherwise, you may spend time refining audio that will need to be generated again.

The AI Act: transparency and permission to use a voice

The European Commission’s Article 50 guidance distinguishes system providers from deployers. It describes machine-readable marking of synthetic output and disclosure of deepfakes to people exposed to them, within the applicable scope. [4]

Assess the system, content and intended use. Information about an audio file’s synthetic origin and permission to use the voice are separate checks: a disclosure does not document authorisation for the campaign.

What to agree before production

A brief for a cloned voice should clarify:

• which recordings will be uploaded and for what purposes, including any further training;
• who may access the model and generate messages;
• the intended clients, products, languages, channels, territories and period of use;
• payment and conditions for expanding the uses;
• who approves scripts and audio versions;
• how access, materials and the model will be managed when the relationship ends.

These points help prepare the agreement; their wording needs to suit the specific project.

For a commercial or jingle, assess the voice within the complete message: clarity of the offer, delivery and coherence with the sonic signature. The ability to generate a new version should be accompanied by a decision about who approves it.

Preparing a commercial, jingle or multilingual campaign? Explore Jingles Factory and bring the scripts, formats and intended uses into the brief.

Questions about AI voices and deepfakes

Is every AI-generated voice a deepfake?

Synthetic audio alone is not sufficient. The AI Act’s definition also requires resemblance to an existing subject and a false appearance of authenticity or truthfulness. [5]

Who must disclose a deepfake audio file’s artificial origin?

Article 50 places this obligation on deployers of systems generating or manipulating content that qualifies as a deepfake, subject to its exceptions. [2]

Does a mark embedded in the file inform listeners sufficiently?

For deployers required to disclose a deepfake’s artificial origin, machine-readable marking alone is insufficient. The Commission requires clear, perceivable information no later than the first exposure to the content. [3]

Sources

Accessed on 7 October 2026.

1. Associated Press, 5 October 2026: Italy’s Meloni follows pop stars in seeking to trademark her voice to combat AI deepfakes.
2. European Commission, AI Act Service Desk: Article 50 — transparency obligations.
3. European Commission: questions and answers on Article 50 transparency obligations.
4. European Commission: guidance on transparency obligations.
5. European Commission, AI Act Service Desk: Article 3 — definitions.