Can someone replicate and freely use anyone's voice using AI?

Question

Suppose person A offers paid voiceover services, and is a popular voiceover artist. Person B wants person A's voice for their revenue-generating internet video projects, because they know it will offer a significant profits-boost. It is legally all right for person B to synthesize person A's voice (without their knowledge or consent) using a neural net and use that audio clip for the videos? That is, are people's voices copyrighted?

Answer 1

Not copyright as such because that is about protecting a 'work' — a voice is not a 'work'. As the court said in one of the following examples, "A voice is not copyrightable. The sounds are not 'fixed.'"

(You could copyright a roar or a yell — some kind of fixed arrangement of sound(s).)

But some jurisdictions have recognised property rights in voices and/or that the voice is protected by the person's 'right of publicity' (the right to control the commercial exploitation of their identity, of which the voice is a part). For example:

Bette Wins Ruling In ‘Sound-Alike’ Lawsuit - AP News

June 23, 1988

SAN FRANCISCO (AP) _ A federal appeals court has reinstated a lawsuit filed by entertainer Bette Midler after an advertising agency allegedly tried to duplicate her voice and singing style in one of its ad campaigns.

The 9th U.S. Circuit Court of Appeals unanimously ruled Wednesday that Midler could pursue her suit against the Ford Motor Co. and the Young & Rubicam advertising agency. The court said certain personal attributes - such as a voice - can be considered property rights, protected by state law.

...

U.S. District Judge Ferdinand Fernandez said Young & Rubicam acted like ″the average thief″ but dismissed Midler’s suit, saying no law prohibits imitation of a singer’s voice.

But the appeals court disagreed.

"A voice is as distinctive and personal as a face,″ the appeals court said. ″When a distinctive voice of a professional singer is widely known and is deliberately imitated in order to sell a product, the sellers have appropriated what is not theirs."

judgment in Midler v Ford

Another case in the US is Waits v Frito-Lay Inc. The US Court of Appeal found that a radio commercial's imitation of the voice of Tom Waits constituted a civil tort, "voice misappropriation".

I'm not aware of any cases involving computer synthesis of voices.

Answer 2

Copyright is the least of B's concerns

Voice, like visual appearance, is an inherent trait of a person. It is not a creative work (let's not dive into philosophy here about creation of human beings). It simply cannot be protected by copyright.

However, that doesn't automatically mean that B can synthesise and publish/make money on A's voice

without their knowledge or consent

— unless B explicitly, prominently and audibly warns their audience that this is not real A's voice and A has actually nothing to do with voicing it.

Such a warning is required as it is difficult to tell AI voice imitation from the real voice owner voicing it, and, by default, the audience will assume it is real A's voice. This is the main B's concern because, unless the warning is given:

• B would effectively be misleading their customers. For many customers (not all of course), the main reason for subscribing/purchasing the media would be the fact that it is real A speaking/signing. If these customers knew it was not the case, they would skip;
• B could make A's voice speak to express views or "testify" facts that A themselves would never pronounce. Such a misuse could damage A's reputation and, hence, be subject of a claim/lawsuit.

Answer 3

"AI" is a function of its training set, which would contain copyrighted data that isn't licensed for that use.

The AI toys we get to see right now are in kind of a legal limbo, because this hasn't been litigated yet, but my initial interpretation is that this is currently merely tolerated by rightsholders, and there is a bit of a rush to create an AI product that is useful enough to cause backlash when it is eventually regulated.

In software development, the use of tools like GitHub Copilot is controversial, because these neural nets often reproduce fragments of the training set, which consists of mostly open source software that has been released under various licenses, some of which stipulate terms under which derived works may be produced.