A voice actor recorded lines for a client. Months later, her voice — not a copy of the recording, but a model trained to speak anything — was for sale on a cloud platform, turning her identity into someone else's inventory. She had never agreed to that, and the law had never been tested on it.
Key takeaways
- The Beijing Internet Court (北京互联网法院) handed down China's first AI-produced-voice personality-rights ruling at first instance on April 23, 2024.
- The plaintiff, a dubber (配音师) identified as Yin, found her voice had been turned into a text-to-speech product and sold through a cloud platform without her consent.
- The court held that a natural person's voice rights (声音权益) extend to AI-produced voice when it is recognizable as that person, and that licensing a recording does not license cloning it.
- Two companies with fault — the firm that recorded and handed over the audio, and the software firm that AI-ized it — were ordered to pay Yin ¥250,000 (≈ US$35,000 / HK$270,000) plus a written apology; three others were cleared of damages.
- In May 2025 the Supreme People's Court (最高法) revisited the issue through its Judicial Answer Net (法答网), reaffirming that using AI-processed voice without permission infringes voice rights.
The chain that turned a voice into a product
The facts read like a small supply chain. Yin, a working dubber, had recorded material for a Beijing culture-and-media company, which held the copyright in the recording. That company passed the audio to a software firm, which used it as training data to build a text-to-speech product. The product was then listed and sold on a Shanghai cloud platform. A Beijing AI company bought access through an intermediary and called the model directly inside its own service.
None of the steps, on their own, involved Yin speaking a new word. The harm, as the court saw it, was that her recognizable voice had been extracted and reproduced as a reusable asset. The culture-media company and the software firm were found to have acted without her permission and with fault; the downstream buyers were judged to lack the requisite fault and escaped damages, though the two upstream firms carried the liability.
What the court actually decided
The ruling rests on the Civil Code (民法典), which protects a natural person's voice by reference to the rules for portrait rights (参照适用肖像权). Voice, the court noted, is distinct, stable, and tied to identity through timbre, pitch, and frequency. The key test is recognizability: if an AI-produced voice lets an ordinary listener connect the sound to a specific person through tone and style, that voice falls within the person's voice rights.
That test is narrower than a full ban on voice cloning. It protects identifiable individuals, not the abstract capability of synthesis. It also draws a line the industry had blurred: holding the copyright to a recording is not the same as owning the right to AI-ize the voice in it. A license to use the audio does not, by itself, authorize turning the speaker into a model.
Why the damages were ¥250,000
The award of ¥250,000 (≈ US$35,000 / HK$270,000) was a discretionary figure, weighed against the infringement's circumstances, the market value of comparable products, and play counts. It was not calculated from a statutory per-use tariff, because none exists for cloned voice in this context. The written apology addresses the dignity side of the harm; the money addresses the commercial side.
The split verdict — two firms liable, three cleared — matters as much as the number. It tells downstream users that buying a model through a normal commercial channel can still be safe if they lacked fault, while the firms that sourced and processed the voice without consent bear the cost. For voice-AI vendors, the lesson is upstream: the legality of the training data, not the cleanliness of the API, is where liability lives.
The 2025 echo from the top court
Roughly a year later, in May 2025, the Supreme People's Court returned to the question through its Judicial Answer Net, a public Q&A channel for judges. The selected answer reaffirmed, in plain terms, that using an AI-processed voice without the person's permission constitutes an infringement of voice rights. Citing the Beijing case as a reference point, the high court signaled the ruling was not a one-off but a reading of where the law sits.
That matters because individual first-instance decisions do not bind the whole country; a Supreme People's Court articulation through an official channel carries more weight as guidance. For the fast-moving voice-cloning market — text-to-speech APIs, dubbing tools, virtual hosts — it closes the "we just bought the model" defense at the sourcing end.
Honest limitations
This article relies on state-media and court-affiliated reporting of the Beijing Internet Court ruling and the May 2025 Supreme People's Court commentary; we did not read the full published judgment text or the Judicial Answer Net original. The companies were anonymized in public reporting as "Company A / Company B" style placeholders, so the exact commercial names and the precise API relationships are not confirmed here. The ¥250,000 figure is reported consistently across multiple outlets but is a court-discretionary amount, not a formula. We have not assessed how this ruling interacts with the September 2025 AI content-labeling measures (which would tag the output but not resolve who owned the voice), nor whether similar cases have since been decided in other courts. This covers one case and one high-court comment, not the full state of voice-cloning law in China.
What readers can do now
- If you are a voice actor or talent, put voice-AI and cloning rights explicitly in every contract — state that recording permission is not model-training permission, and that any AI use needs separate written consent.
- If you build or buy voice-AI, trace the training data to the speaker's consent, not just to the vendor's terms; downstream "we bought it legally" is no shield if the source lacked permission.
- If you publish AI-voiced content, pair it with a label and credit, and avoid mimicking a specific identifiable person without authorization — recognizability, not intent, is what the court tests.
