A voice actor records a script, gets paid, moves on. Months later his voice is for sale on an app — cloned, packaged, and pitched to strangers — and he never agreed to any of it. That quiet theft is now a test case for how China regulates synthetic media.
The line between a useful text-to-speech tool and a stolen identity turned out to be thinner than the industry expected. China's answer arrives on two tracks: a court ruling that made cloned voices someone's property, and a national labeling rule that makes hidden synthesis illegal to hide.
A voice actor's voice went to market without him
In April 2024, the Beijing Internet Court handed down what it called the country's first AI-voice infringement case. The plaintiff, a voice actor identified as Yin, discovered that his recorded voice had been turned into a text-to-speech product and sold on an app called "Moyin Gongfang" under the name "Mo Xiaoxuan."
The court found that a culture-media company and a software company had used Yin's recordings to build the clone without his consent, and that owning the copyright to a recording does not mean owning the right to AI-clone the voice inside it. The ruling ordered a written apology and ¥250,000 (≈ US$34,000 / HK$270,000) in damages.
The legal hook is the Civil Code, which protects a natural person's voice by reference to portrait-right rules: if the public can recognize you by your tone, that voice is yours to control.
The new labeling rulebook
The bigger shift came from regulators. In March 2025, four ministries — the Cyberspace Administration of China (CAC), the Ministry of Industry and Information Technology, the Ministry of Public Security, and the National Radio and Television Administration — jointly issued the Measures on Labeling of AI-Generated and Synthetic Content.
The rules took effect on September 1, 2025. They cover text, images, audio, video, and virtual scenes produced with AI, and they demand two kinds of marks:
- Explicit labels — a visible or audible tag a user can actually perceive, like a watermark, a caption, or an audio cue.
- Implicit labels — machine-readable metadata baked into the file, so platforms can detect synthetic content as it travels.
For audio specifically, the measures say a provider must add a voice prompt or a rhythmic cue at the start or a suitable point in the clip, and must embed metadata in the file itself.
What counts as a label
The measures are practical about where the tag goes. A text generator should flag its output in the text or the interface. An image tool should mark the picture. A video tool should mark the frame. An audio tool should mark the sound.
Services that let users download or share synthetic content must make sure the explicit label travels with the file. Platforms that republish it must check for the hidden label and, if they can't verify one, add their own notice that the content is synthetic. Deleting or tampering with a required label is explicitly forbidden.
According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, voice is the next front in synthetic-media regulation because it is cheaper to clone and harder to flag than video.
Why voice is the hardest deepfake to spot
Cloned voices now power everything from customer-service bots to microdrama (短剧) dubbing and short-video narration. A voice needs only a few seconds of sample audio to imitate, and the result often passes a casual ear check that a face swap would fail.
That makes labeling a weaker shield than enforcement. The Beijing case matters precisely because it attacked the business model, not just the watermark: it told companies that selling a clone without a license is a payable offense, not a gray area. The National Radio and Television Administration had already, in late 2024, warned platforms against AI "remakes" that distort original works without clear marking.
How labels travel across platforms
The measures push responsibility up the chain, not just onto the generator. A tool that creates the clip must label it; a platform that republishes it must check for the hidden tag and, if none survives, attach its own notice. That chain-of-custody logic matters because synthetic audio rarely stays where it was born — a cloned voice clip jumps from one app to a messaging group to a short-video feed within minutes. Making the label self-describing metadata, rather than a visible stamp a user can crop out, is what lets the signal survive that journey. The unresolved piece is cross-border: a clip labeled inside China carries no enforceable meaning once it lands on an overseas service, so the rule's real reach is domestic platforms and the domestic app stores that distribute the generators. That is why the court precedent and the labeling rule are complementary — one defines the right, the other enforces the marking, and neither fully closes the gap for content that crosses borders.
Honest limitations
- The ¥250,000 award is a single first-instance ruling, not a statutory damages floor; later cases will set the real range.
- The labeling measures tell providers what to mark, but enforcement against small app stores and overseas tools remains uneven.
- "Recognizability" is the legal test for voice theft — a clone that the public cannot link back to a specific person may fall outside the protection, which leaves edge cases open.
What readers can do now
- Check the audio cue — when you hear suspicious narration in ads or shorts, look for a stated AI label before trusting the source.
- Protect your own voice — avoid handing raw voice samples to unlicensed apps; a few seconds is all a clone needs.
- Report, don't assume — if a clone uses a known voice without consent, Chinese courts have now shown the claim is winnable.