Beijing Internet Court: AI-Processed Voices Protected Under Natural Person’s Voice Rights（Judgment issued on April 23, 2024）

Mengchun’s Comments

Advances in AI deep synthesis technology now allow voices to be generated and processed by AI. At the same time, unauthorized use of natural persons’ voices for AI training has increasingly led to legal disputes. In this context, the court ruling confirms that voice rights can extend to AI-generated voices that remain recognizable, and it establishes a framework for assessing recognizability.

Key Issue

Can a natural person’s voice rights extend to AI-generated voices? Specifically, how can one determine whether a synthesized voice remains recognizable as originating from a specific individual?

Analysis

Voice rights are protected only if the voice is recognizable. Each person’s voice is unique due to differences in vocal characteristics, pronunciation habits, vocal frequency, and the physical structure of vocal organs, as well as factors such as age and personality. This uniqueness allows listeners to associate a voice with a particular individual and thus establish identity recognition.

Whether an AI-processed voice falls within the scope of a natural person’s voice rights depends on whether it still allows identification of the individual. Courts assess this using three criteria: subjective standard, objective standard, and mode of use. Two scenarios are analyzed:

1. Small-scale use: AI is trained on voices from one or two individuals.
   • Objective standard: voiceprint recognition or confirmation
   • Usage comparison: timbre, intonation, pronunciation style compared with the natural voice
   • Subjective standard: social recognition of the individual
     • Highly recognizable figures → general public should be able to identify them
     • Professionally known individuals (e.g., voice actors) → recognition assessed within the relevant audience
     • Ordinary individuals → lacking notable recognizability; only objective measures apply
2. Large-scale use: AI is trained on dozens or hundreds of voices.
   • The resulting AI-generated voice is generally unrecognizable.
   • Voice rights do not extend to such AI-generated voices.
   • Unauthorized use may still constitute infringement if verified by objective standards.

Case Background

The plaintiff, Yin Mouzhen, is a voice actor. In May 2023, he discovered that works using his recorded voice were widely distributed across multiple popular apps. Investigation traced the voices to a text-to-speech product operated by Defendant One, a technology company.

• In 2019, the plaintiff recorded audio for Defendant Two, a media company. Defendant Two held the copyright to the recordings.
• On May 10, 2019, Defendant Two authorized Defendant Three, a software company, via a Data Authorization Agreement to use the recordings, using a non-signed authorization form.
• Defendant Three used only one of the plaintiff’s recordings to create the text-to-speech AI product. Defendant One and Defendant Five purchased and deployed the AI product on their platform, which was streamed over 3.2 billion times.

The plaintiff sued all five defendants, alleging severe infringement of voice rights. The defendants denied liability, arguing that AI-generated voices differ from natural voices and are marked with watermarks, breaking any connection to the plaintiff’s persona.

Judgment Summary

The court referenced the Civil Code and analogized voice rights to portrait rights. Portraits reflect identifiable external images; similarly, voices possess distinctiveness, uniqueness, and stability, allowing the audience to associate them with the individual. For professional voice actors, if the audience can associate the voice with the person, recognizability is established.

The court held that AI-processed voices fall within the protection scope if the public can identify the individual based on timbre, intonation, and pronunciation style. In this case, the AI-generated voices closely matched the plaintiff’s original voice, and listeners could associate them with him.

Outcome:

• Defendants Two and Three, who AI-processed and used the plaintiff’s recordings without permission, infringed on the plaintiff’s voice rights.
• They were ordered to apologize and jointly compensate the plaintiff for economic losses of RMB 250,000.
• Defendants One, Four, and Five were not held liable as they lacked subjective fault.
• No appeals were filed; the judgment is now final.

Court:
Beijing Internet Court

Final Decision Date:
April 23, 2024

Case No.:
(2023)京0491民初12142号
(2023) Jing 0491 Min Chu No. 12142