Does Generating 'Voices' with AI Constitute Copyright Infringement? (#1 Development and Training Phase)
With the advancement of generative AI, it has become possible to easily learn and generate the ‘voices’ of existing singers and voice actors. In the business scene, this technology allows for the learning and creation of new ‘voices’ in app development, game creation, and anime production.
Having generative AI learn and create new ‘voices’ based on those of existing singers and voice actors could potentially constitute copyright infringement or other illegal activities under Japanese law.
In reality, there is currently no clear interpretation regarding these issues. What exactly are the legal rights associated with a ‘voice,’ and under what circumstances could it become a problem under the Japanese Copyright Act?
In this series, we will explain this issue in two parts, considering specific usage patterns. This article, the first part, discusses potential rights infringements that may occur during the development and learning stages of generative AI. Legal issues at the generation and usage stages are explained in this article (Part 2: Generation and Usage Stages). Please refer to it as well.
Three Legal Rights Surrounding a Person’s “Voice” Under Japanese Law
What legal rights are associated with a person’s “voice”? To address this issue, it is necessary to consider two perspectives regarding the “voice”:
- What the voice is saying, and
- The sound personalityistics of the voice.
In other words, the first point concerns the “content” of the voice, while the second point relates to the “sound” of the voice.
For example, if different voice actors perform the same line “Good morning,” the content in the first case remains the same, but the sound in the second case differs.
Based on these perspectives, under current Japanese law, there are three legal rights that can arise in relation to a person’s “voice”:
| ① Copyright | May arise from the voice’s “content” |
| ② Neighboring Rights (limited to the rights of performers) | May arise from both the voice’s “content” and “sound” |
| ③ Publicity Rights | May arise from the voice’s “sound” |
About Copyrights Under Japanese Law
Copyrights arise when the “content” of a voice qualifies as a work of authorship.
For instance, when reading aloud a famous novel, the voice may be subject to copyright. However, it is important to note that in such cases, the copyright holder is the author of the novel, not the “voice owner = speaker.” Therefore, if you create a synthesized voice reading the content of a famous novel using generative AI, this act could potentially infringe upon the copyright of the novel’s author.
Conversely, if the content of the voice consists of mundane everyday conversations of ordinary people, no copyright will arise. This is because commonplace daily conversations do not qualify as works of authorship and are not protected under copyright law.
About Neighboring Rights Under Japanese Law
Neighboring rights (limited to the rights of performers) can arise when the content of a voice is considered a work, such as when the voice accompanies a form of expression like reading aloud.
As mentioned in the section on copyright, when a voice performs a “reading,” it is considered a “performance,” which may give rise to neighboring rights for the reader. Unlike the case of copyright, it is important to note that the holder of the neighboring rights is not the author of the novel but the reader who is actually vocalizing.
Understanding Publicity Rights in Japan
Publicity rights, as recognized by Japanese case law (Supreme Court decision on February 2, 2012 (H24.2.2)), refer to the exclusive right to utilize the customer-attracting power of a person’s name, likeness, and other personal attributes.
| ▶︎ Supreme Court decision on February 2, 2012 (H24.2.2) (The Pink Lady Case) ■ Key Points of the Decision ① The use of a name, likeness, etc., as an object of appreciation in itself, as in goods, ② The attachment of a name, likeness, etc., to goods with the purpose of differentiating the goods, ③ The use of a name, likeness, etc., as advertising for goods, where the sole purpose is to exploit the customer-attracting power of the name, likeness, etc., constitutes an infringement of publicity rights and is illegal under tort law. ■ Commentary by the Investigating Officer (Supreme Court Case Commentary, Civil Matters, 2012 (H24) Volume 1, page 18) The term “likeness, etc.” in the three categories of this decision refers to personal identification information of the individual, which includes, for example, signatures, autographs, voices, pen names, stage names, etc. |
According to the Pink Lady Case, there is room for the emergence of publicity rights even in one’s voice. If it is determined that the voice belongs to a person with customer-attracting power, such as a real voice actor, actor, or singer, publicity rights arise regardless of the ‘content’ of the voice. Furthermore, if the voice is used in any of the three infringement patterns indicated in the Pink Lady Case, it constitutes an infringement of publicity rights.
Three Usage Patterns in the Development and Learning Phase
Saying “generating voices with generative AI” may sound straightforward, but the process can be divided into the following two stages:
- Development and learning phase
- Generation and utilization phase
The first stage is carried out by AI developers, while the second is performed by AI users.
When visualized, these processes are as follows:
During the development and learning phase, developers collect and accumulate original human voice data as training data for AI development, creating a training dataset. This dataset is then fed into the AI for machine learning, resulting in a trained model. On the other hand, in the generation and utilization phase, the original data is input into the generative AI that has completed machine learning to produce and utilize AI-generated content.
The following three patterns can be anticipated as usage patterns during the development and learning phase:
- Pattern 1: Collecting, accumulating, processing, and using human voice data as training data for AI development
- Pattern 2: Selling or publishing the training datasets used in AI development
- Pattern 3: Selling or publishing the generative AI itself
Below, we will briefly explain the potential for rights infringement in each of these usage patterns.
Pattern 1: Collecting, Storing, Processing, and Utilizing Human Voice Data for AI Development
First, we will explain potential rights infringements that may occur during the stages of collecting, storing, processing, and utilizing human voice data for training AI.
Relation to Copyright Law
The use of human voice data in Pattern 1 specifically refers to the act of developing the AI itself. AI development falls under the category of “information analysis” as per Article 30-4, Paragraph 2 of the Copyright Law (hereinafter referred to simply as “the Law”), and therefore, the use of copyrighted works necessary for such development is not considered an infringement of copyright (Article 30-4).
However, there is a significant exception to the above. If the purpose of creating the training dataset is to generate AI products that possess the essential personalityistics of the original data’s expression (expression output purpose), then Article 30-4 does not apply, and the use becomes illegal.
In other words, if the use of voice data from other voice actors aims to replicate or pay homage to the distinctive voice of a specific voice actor, there is a possibility that such use could be considered an infringement of copyright due to the expression output purpose.
Relation to Neighboring Rights
In relation to neighboring rights, Article 102 applies the provisions of Article 30-4 to neighboring rights as well, so the act of performing for the development of generative AI does not constitute an infringement of neighboring rights in principle.
Relation to Publicity Rights
The issue of publicity rights arises when the development of generative AI aims to create the ‘voice’ of a specific, well-known individual with customer appeal.
Whether or not this constitutes an infringement of publicity rights can be informed by the three types of infringement modes in the aforementioned Pink Lady case.
Firstly, the act of developing generative AI with the purpose of creating the ‘voice’ of a specific celebrity does not correspond to the infringement modes in the Pink Lady case. However, if the act is “solely aimed at utilizing the customer appeal of names, portraits, etc.,” it may constitute an infringement of publicity rights and thus be an unlawful act.
For the utilization of customer appeal to occur, it is necessary for third parties to perceive that the voice belongs to the said celebrity during the development of the generative AI, particularly when creating the training dataset. If third-party customers do not perceive it, customer attraction simply cannot occur. However, there is usually no room for third-party customers to be involved in the development stage of generative AI.
Therefore, it can be said that there is little room for such use to constitute an infringement of publicity rights.
Pattern 2: Sale and Publication of Training Datasets Used in AI Development
This section explains potential rights infringements that may occur during the sale and publication of AI training datasets.
Relation to Copyright
When the original data is stored in the training dataset in its original form or slightly processed, the act of selling or publishing the training dataset may constitute an infringement of the transfer rights (Article 26-2) or public transmission rights (Article 23) of the copyrighted work or derivative work (Article 28). Therefore, performing these actions without the consent of the copyright holder will result in copyright infringement.
However, as with the above, Article 30-4 stipulates that “for the purpose of information analysis,” it is permissible “to use by any means necessary within the recognized limits.” Thus, the transfer or publication for the development of generative AI does not constitute copyright infringement as long as it is within the necessary limits.
Relation to Neighboring Rights
Similarly, as per Article 102, the provisions of Article 30-4 regarding copyright are applied accordingly, so the sale and publication of training datasets for the development of generative AI do not generally constitute an infringement of neighboring rights.
Relation to Publicity Rights
Some training datasets contain the voices of specific celebrities in a format that can be reproduced as is. However, training datasets are typically used solely for the development of generative AI and do not fall under the category of “using the name, likeness, etc., as an independent object of appreciation in goods,” as indicated in the Supreme Court’s decision in the Pink Lady case.
Therefore, it can be said that there is little room for the act of using such datasets to infringe upon publicity rights.
Pattern 3: Sale and Publication of the AI Itself
Here, we will explain potential rights infringements that may occur at the stage of selling or publishing the trained AI model itself.
Relation to Copyright
Unlike the training datasets, it is not conceivable that the trained AI model retains any creative aspects of the original data (copyrighted works). Therefore, the AI itself, that is, the trained model, cannot be considered a derivative work of the original data, and its publication or sale does not constitute copyright infringement.
Relation to Neighboring Rights
As with copyright, since it is not conceivable that the trained model retains any creative aspects of the original data, the sale and publication of the AI itself do not infringe upon neighboring rights.
Relation to Publicity Rights
Even an AI capable of freely and accurately generating the voice of a specific celebrity does not fall under the three types of infringement indicated by the Supreme Court’s decision in the Pink Lady case. However, such AI is typically marketed on the value of being able to generate the voice of a specific celebrity with high precision, and customers usually purchase the AI for its ability to do so. Therefore, the sale of such AI could likely be considered an act similar to the three types of infringement and potentially constitute a violation of publicity rights.
Conclusion: Consult Experts on the Relationship Between Generative AI and Copyright
Thus far, we have explained the legal rights associated with human voices and the problematic actions that may arise when utilizing them, based on specific examples.
It is crucial to consider the legal rights of human voices by separating ‘content’ and ‘sound’, and to understand the potential applicability of copyright, related rights, and publicity rights. While this article has focused on the development and learning stages as the first part, the second part will delve into the generative and utilization stages.
Related article: Could Generating Voices with AI Lead to Copyright Infringement? (#2 Generative & Utilization Stages)
Guidance on Measures Provided by Our Firm
Monolith Law Office is a law firm with extensive experience in both IT, particularly the internet, and legal matters. In recent years, generative AI and intellectual property rights surrounding copyright have garnered significant attention, and the necessity for legal checks has been increasingly on the rise. Our firm provides solutions related to intellectual property, which are detailed in the article below.
Areas of practice at Monolith Law Office: IT & Intellectual Property Legal Services for Various Companies
Category: IT
