A voice can now be cloned in seconds. Why consent and clear licensing terms must come first, and what an on-chain record of rights can and can't do.
A singer's voice is the most personal thing an artist brings to a record: years of practice and a tone listeners recognize before the first word is finished. Generative AI can now reproduce much of that from a small amount of audio. The question is no longer whether that is possible. It is who gets to say yes, on what terms, and how anyone can tell afterward.
In recent weeks, courts, lawmakers, labels and AI companies have all given partial answers. Taken together, they point one way: consent comes first, and it comes with terms attached.
A court says it plainly
On September 29, the Shanghai No. 1 Intermediate People's Court upheld a ruling ordering a company to pay 50,000 yuan in damages to a voice actor whose voice had been cloned with AI, as reported by the Seoul Economic Daily and in the court's own account carried by Chinese media. A forensic comparison found that 90% of the AI-generated voice was similar or highly similar to the actor's own.
In the Seoul Economic Daily's translation, the court held that "using a person's voice as training data without their consent and imitating their tone, intonation and pronunciation to create a voice by which a specific individual can be identified constitutes a violation of voice rights." The company, which controlled the training data, could not show that its source material was lawful, and that counted against it.
The court treated the voice as part of the person, not a by-product of a recording. And the burden fell partly on the company to show where its data came from. When nobody can say what went into a model, that silence has a cost.
The law elsewhere is still a patchwork
In Japan, a review committee set up by the Ministry of Justice published interpretive guidelines on August 7 that confirm a person's voice falls within what publicity rights protect. Under those guidelines, generating a voice that closely resembles a specific singer and publishing it for revenue can infringe those rights. The report is not legally binding, but it is a significant reference point.
In the United States, protection for voices still depends heavily on state law. In Lehrman v. Lovo, decided in July 2025, a federal court in New York dismissed most of two voice actors' federal trademark and copyright claims over alleged AI clones of their voices. Copyright, the court wrote, "does not extend to something as abstract and intangible as a 'voice.'" What survived were claims under New York's Civil Rights Law, which bars using a living person's voice for advertising or trade "without having first obtained the written consent of such person." Which state's law applies can decide whether a cloned voice has any remedy at all.
A federal answer is on the table but not yet in force. The NO FAKES Act of 2026 would create a federal right to authorize digital replicas of a person's voice and visual likeness, licensable during the person's lifetime. It cleared the Senate Judiciary Committee in June. On September 30, an attempt to pass it by unanimous consent drew an objection on the Senate floor, and the bill is stalled until at least the post-election session.
In many places, the strongest protection an artist has today is not a statute. It is a contract.
Consent is becoming the price of entry
The industry seems to be reaching the same conclusion.
On September 10, Universal Music Group and ElevenLabs announced a multi-year licensing agreement that begins with a platform, still in development, built on "licensed music and artist participation," including what the companies call "personalized vocal experiences." Participation is limited to artists and songwriters who opt in. No launch date, artist list or payment terms have been published.
Litigation is pushing in the same direction. In a new complaint filed on September 18, Universal and Sony allege that Suno copied 60,202 of their recordings without a license to build its models. The labels argue that Suno's own licensing deals "confirm that a functioning market exists for licensing sound recordings to train generative AI models." These are allegations, not findings, and Suno argues that training on copyrighted material is fair use. But permission is being treated as something with a market and a price.
Meanwhile, three AI voice companies that launched new models in early October all used the word "consent" in their announcements, as Voice Over Herald noted on October 18. The same piece pointed out what those announcements left out: how the people behind the voices were paid, how long the license runs, and what happens if they want out.
That is the gap we care about. "We have consent" is a claim. A license is a set of terms.
What a real voice license should answer
For any artist asked to license a voice, or a voice model built from their recordings, the useful questions are concrete:
- Scope. Which products, which clients, which languages, which territories?
- Training versus output. Is the permission to train a model, to generate new performances, or both? These are different uses and deserve separate answers.
- Term and withdrawal. How long does it last, and what actually stops if the artist withdraws?
- Approval. Does the artist see and approve uses before release?
- Pay. A one-time fee, or payment each time the voice is used?
- Endings. If the service shuts down or is sold, what happens to the voice model and the source recordings?
None of this is exotic; the music business already applies it to sync licenses. What is new is that one recording session can now stand in for performances nobody has imagined yet.
Where on-chain records fit, and where they don't
This is where rights recorded on a blockchain can help, if we are clear about what they are.
An on-chain record can document, with a timestamp that can't be quietly edited later, that a specific artist granted a specific permission on specific terms: this voice, for this use, for this period, with this split. It can also document when that permission was changed or withdrawn. That makes it a record of consent and licensing: a shared reference for artists, labels, platforms and, if needed, courts.
What it cannot do matters just as much. A record does not stop someone from cloning a voice without asking. It does not decide whether a use is lawful; that is still the job of contracts, statutes and courts, and as the cases above show, the law differs from place to place. A record of rights is only as good as the agreement behind it. What it changes is the question a company has to answer when something goes wrong. Instead of "can you prove you never agreed?", the question becomes "where is the record that you did?"
The Shanghai ruling shows why that shift matters. A company that could not account for its training data paid for it.
Where XECHO stands
XECHO is the music platform we are building. It is in invite-only beta, with access by request. XECHO is designed to record rights on-chain and handle royalties through smart contracts. In the current beta, both run on testnet only.
We don't think AI is the enemy of artists. Some will choose to license their voices. That should be their choice, made on clear terms and recorded in a way that can be checked later.
Your voice is not training data unless you say it is.
XECHO is in invite-only beta. Access is by request at xecho.pro, and our team reviews each request.
