“Don’t steal our voices”: the dubbing world doesn’t want AI

“Don’t steal our voices”: “Don’t steal our voices”. This is the call launched through the initiative of the same name, led by 21 unions and associations of voice professionals in cinema, advertising or video games. Gathered in an international coalition called UVA, for United Voice Artists, they represent seven European countries, including France, but also the United States, Turkey and South America. All warn against voices generated by artificial intelligence in a open letter posted Thursday, May 25.

« Artificial intelligence must be a tool at the service of man, and not something that comes to steal our profession, our creations, our interpretation, and above all our personalities, because the voice is part of our personality like our face “, explains Patrick Kuban to Numerama. The professional voice actor and co-founder of the collective The voices, who signed the platform, is notably the voice of the Canal+ channel for trailers, of RTL 2 radio (Le Son Pop-Rock) and also works on commercials for brands and documentaries. He agreed to answer our questions.

What voices to train AI?

As with generative AIs that generate images (Midjourney) or text (ChatGPT), companies that develop voice-based artificial intelligences must rely on a large volume of data to develop their models. And this is also the starting point of the initiative.

In April, French members of Les Voix reported strange recording contracts ” for research purposes ”… which in fact ceded rights to Voiseed, a Milanese start-up specializing in artificial intelligence. Patrick Kuban analyzes: “ When you say on a contract that it is for research purposes, you fail to specify that these voices are in fact going to be used to produce a result and be marketed in a system “, protests the professional, who is campaigning for this use to be subject to prior authorization.

Many companies have specialized in AI-generated voices, such as Vall-E, Lyrebird or ElevenLabs. In 2021, the horror film Every Time I Die has for example been dubbed into Portuguese and Spanish thanks to AI, by the Israeli start-up DeepHub.

While some companies would use studio voices as training material, according to Patrick Kuban, others ” draw on the Internet on works that are protected by copyright, and in particular audio books, films and series, extracts from radio programs, podcasts, etc. ».

This question of the massive use of content not free of rights to train AIs is at the heart of a legal puzzle. The IA Act, the proposed regulation of artificial intelligence by the European Union, provides for its part to impose that all generative AIs disclose the content protected by copyrights that they have used to train.

The voice, a sensitive biometric data

According to the voice professional, the problem is twofold: “data mining, the right to search, is used in an illicit way, since there is a commercial utility, but there is also the presence on these databases of entrainment of sensitive biometric data which is the voice, which is mentioned in the General Data Protection Regulation (GDPR) : its use being subject to the authorization of the persons ».

A question that is all the more sensitive since malicious actors are already using vocal imitations of real people. These could even pose national security concerns, the voice actor points out. ” You imagine a minister whose voice would be used in a conversation that would be stolen in a chain of command. It would be dramatic “, he imagines.

…even in the hereafter

What about the artists who have indeed granted their authorization? This is the case of the American James Earl Jones, who lends his voice to Darth Vader for forty years. Last September, the 92-year-old comedian sold a sound bank of his robotic bass voice and signature breath to a Ukrainian company, Respeecher. Even after his death, his voice clone will be able to continue to double Vader in future Star Wars.

“There is an ethical point of view, to say to oneself should the dead actors speak or not, should the rights holder agree or not “Once the person has died, believes Patrick Kuban, for whom it is still too early to ask the question.

“This is not a voice”

Finally, it also raises the question of the mandatory warning of content generated by IA, wanted by the IA Act, and recommended by France for images and texts. Technically, it could go through “ audible tags when you listen to a podcast, an advertisement or audio book for example, which tell the listener “What you are listening to is generated by a robot and a synthetic voice” “Or by a visual banner before a video, supposes Patrick Kuban.

To trace the artificial nature of this content, he imagines a vocal “watermark” system. “If generative AI systems are certified with GDPR databases, everything that comes out must be “watermarké” in the audio signal. We can add a signal that cannot be heard, a barcode in the voice, which will link the voice we hear to a contract “, proposes the co-founder of the Voices.

Profession in danger, applications already on the market

For Patrick Kuban, the danger is real for the profession. Dubbing actors are an essential link in the sector: “ If you break the chain, everything collapses: if there are no more actors in the studio, there is no more recording studio, sound engineers, authors ».

And the news would almost prove him right. In February, the production agency Prodigious, for example, announced that it was developing an application generating synthetic voices via artificial intelligence, called TalkBox. Able to speak 83 languages ​​and different accents (up to 14 accents in English alone) and imitate male, female and children’s voices. The application initially targets minimal interventions such as legal notices on television and radio, but eventually it could concern sponsored content, demonstration videos or even documentaries.

---

Do you want to know everything about the mobility of tomorrow, from electric cars to pedelecs? Subscribe now to our Watt Else newsletter!