You’re producing a podcast, an advertising campaign, or an audio guide. The voice sounds familiar, professional, perfectly aligned with the brand. There’s just one thing: the person behind that voice never set foot in the studio. Voice cloning makes exactly this possible — and it raises questions that most standard contract clauses are nowhere near equipped to answer.
This is no longer an emerging technology. Language models can clone a voice from just a few minutes of audio. For producers, agencies, and media companies, the result is a genuine legal grey area: Who has the right to clone a voice? Who owns the output? And what happens when a training dataset contains recordings that were never properly licensed? This article maps out the key dimensions and identifies where contracts need to be updated — now.
The Voice as a Personality Right: What German Law Protects
German law does not recognize an explicit “right to one’s voice” as a codified legal concept. What it does recognize is the general right of personality under Article 2(1) in conjunction with Article 1(1) of the Basic Law (Grundgesetz). Over the decades, courts have derived from this right a form of protection for the voice as a distinctly personal characteristic.
A voice is identifiable, distinctive, and inseparable from the individual. Using it commercially without consent constitutes an infringement of personality rights. This applies to audio recordings — but also to synthetic reproductions that imitate a real person’s voice. Whether an AI-generated voice clone qualifies legally as a “recording” or as an independent creation has not yet been decided by Germany’s highest courts. The prevailing view in German legal scholarship, however, points toward similarity as the determining factor: if a cloned voice recognizably replicates a real person, personality rights protections apply.
For production teams, the practical implication is clear: consent to record is not consent to clone. Anyone commissioning voice talent for audio formats will need an explicit clause going forward — one that specifies whether, and under what conditions, a voice may be used for training data or synthetic reproduction.
Dataset Provenance: The Liability Question Nobody Sees Coming
Behind every voice cloning tool lies a training dataset. And this is precisely where many production professionals are still underestimating the risk: where did the audio data used to train the model actually come from?
A pattern has emerged from several high-profile cases in the United States. Voice actors and musicians have discovered that their publicly available recordings — from podcasts, audiobooks, and YouTube videos — were incorporated into training datasets without any licensing agreement. On June 24, 2024, Universal Music Group, Sony Music, and Warner Records, through the RIAA, filed landmark lawsuits against AI music services Suno and Udio, alleging large-scale copyright infringement. (Recording Industry Association of America) In the EU, the AI Act — which has been entering into force in stages since August 2024 — significantly raises the bar for transparency and documentation of training data, particularly for high-risk systems and foundation models.
For agencies and media companies using commercial voice cloning services, this creates a genuine due diligence obligation. Deploying a tool without scrutinizing its data foundation means accepting potential co-liability if the model turns out to be built on unlicensed recordings. The question “Can I demonstrate where the training data came from?” is fast becoming a contractual matter: providers of voice cloning services should be expected to offer warranties on dataset provenance — much as software vendors today make representations about open-source license compliance.
Where Standard Contracts Fall Short
A typical voice talent contract covers fees, usage rights by medium and territory, and an exclusivity period. What it does not typically cover: synthetic reproduction, cloning the voice for future productions, use as training data, or AI-assisted post-processing.
This isn’t an oversight — these clauses simply didn’t exist because the technology didn’t exist. Now it does. And the silence in the contract has become a liability, in both directions. Producers who assume a broad usage rights clause covers voice cloning may find themselves on shaky legal ground. Voice talent who assume their voice is automatically protected may discover too late that they never negotiated an explicit prohibition.
At a minimum, contracts today should address the following:
Cloning permission or prohibition: An explicit statement on whether the voice may be used for synthetic reproduction — and if so, under what conditions (single project, time-limited, with separate compensation).
Training data exclusion: A clause prohibiting the use of recordings from the project as input for AI model training, unless a separate agreement is in place.
Disclosure obligation for AI use: A requirement for the commissioning party to disclose when AI tools are used in any stage of voice processing.
Compensation model for synthetic use: If cloning is agreed upon, it requires its own licensing structure — not a blanket buyout bundled into the studio session fee.
Deletion obligations and data control: Who retains access to raw recordings after project completion? When are they deleted?
These are not considerations reserved for large-scale productions. They represent baseline practice for any agency commissioning voice work today.
The EU AI Act as a New Framework for Production Practice
The AI Act introduces a degree of regulatory clarity into a debate that has, until now, been shaped largely by individual cases and national case law. For voice cloning, two areas are particularly relevant: mandatory labeling of AI-generated content, and requirements for providers of foundation models.
Article 50 of the AI Act requires providers of AI systems that generate synthetic audio content to label that output accordingly. This includes voice cloning. Anyone deploying a cloned voice in an advertising campaign or media format will need to make clear that the audio is AI-generated. The precise implementation is still being defined through secondary legislation, but the direction is unambiguous.
For production contracts, this translates into a concrete obligation: labeling requirements become a contractual responsibility. Commissioning parties must ensure that their contractors are aware of these requirements and comply with them. Failing to address this contractually means carrying the liability yourself. (EUR-Lex)
New Compensation Models: What the Industry Can Learn
The music industry has already provided a blueprint — a painful one, but instructive. When streaming upended the economics of music licensing, it took years for royalty models to emerge that treated creators fairly. Voice cloning is at a comparable inflection point, with considerably less time to adapt.
Some voice talent agencies and unions — primarily in the United States — are already developing frameworks that treat synthetic use as a distinct licensing tier. In its 2023 TV/Theatrical contracts and 2024 TV Animation agreements, SAG-AFTRA secured notification, consent, compensation requirements, and bargaining rights for the creation and use of synthetic voices and digital replicas. (SAG-AFTRA) The underlying principle is straightforward: cloning a voice is not the same as recording it once. It creates a durable asset that can be deployed repeatedly. That distinction needs to be reflected in how talent is compensated.
In the German-speaking market, comparable collective bargaining structures are largely absent. Some agencies address this contractually on a case-by-case basis. An industry-wide solution is not yet on the horizon. That gap is worth understanding — even from the commissioning side. Offering fair terms builds more durable working relationships with voice talent than relying on contractual silence.
What to Do Now: Practical Steps for Producers
Voice cloning is not a future problem. It is a problem for your next production. Anyone currently commissioning or producing voice recordings should take three concrete steps:
Review existing contract templates. Do they include explicit provisions on synthetic reproduction? If not, they need to be updated before the next project gets underway.
Ask hard questions of voice cloning tool providers. What training data was used? Is there dataset documentation? Are warranties on the lawfulness of the data available? If you don’t get answers, treat that as a signal.
Build labeling compliance into your workflow. Even where the AI Act is not yet fully enforceable: organizations that establish transparent processes for identifying AI-generated voice content now will be well-positioned when the obligations come into full effect.
The right to one’s voice is not an abstract concern for media law specialists. It is a production issue — one that plays out in contract clauses, tool selection, and compensation structures. The technology has opened a new chapter. The contracts need to catch up.
Sources
Labeling requirements for synthetic audio content under Article 50 of the AI Act — EUR-Lex
Collective bargaining provisions for synthetic voice use in the United States — Los Angeles Times, November 13, 2023
