False note from OpenAI | AI voice generation in the spotlight

False note from OpenAI | AI voice generation in the spotlight
False note from OpenAI | AI voice generation in the spotlight

(Paris) How are voices generated in artificial intelligence (AI) applications? The question is burning after American actress Scarlett Johansson accused the company OpenAI of having copied her stamp without her consent for ChatGPT.


Published at 8:30 a.m.

Joseph BOYLE

France Media Agency

Sam Altman, managing director of the group, apologized to the actress last week and announced the suspension of Sky, the offending voice.

Approached by the company last September, Scarlett Johansson, who lent her voice to an artificial intelligence system in the film Her about ten years ago, refused to allow it to be used in the new version of ChatGPT.

OpenAI, for its part, denied having imitated the voice of Mme Johansson, claiming to have developed her tool with that of another professional actress.

Hyper-realistic audio tricks, risk of fraud, disinformation: the ability of AI systems to imitate human voices regularly hits the headlines.

Last year, the start-up Eleven Labs, which is developing a voice cloning tool using artificial intelligence, had warned against the abusive use of its software.

Users of the anonymous forum 4Chan had notably shared messages imitating the voices of celebrities to make them pronounce texts of a racist, sexist and homophobic nature.

In one of them, a “fake” Emma Watson was reading a passage from “Mein Kampf”.

“Paradigm shift”

The technology developed largely thanks to an open-Source program called Tortoise, launched two years ago, said Danish entrepreneur Victor Riparbelli, director of the start-up Synthesis.

His company transforms text into a video with an avatar generated by AI. It hires actors whose voice and appearance are the subject of a two-year contract, with an option to renew, explains Mr. Riparbelli, met on the sidelines of the VivaTech new technologies show in Paris last week.

The Tortoise machine learning program analyzed thousands of hours of audio recordings and served as the basis for Eleven Labs’ application.

“It was an important paradigm shift,” continues Mr. Riparbelli.

OpenAI uses similar programs, although it does not disclose the details.

With ChatGPT 4.0, a user can whisper a few sentences in French to the application which will be able in a few minutes to reproduce this voice and use it to narrate a short video in five different languages, as OpenAI showed during of a demonstration in Paris.

As the start-up American, hundreds of companies now offer voice cloning, which is becoming more and more realistic and nuanced.

Among them, TALKR.ai is a French publisher of virtual voice assistants created by artificial intelligence.

This service could handle 25% to 30% of customer service calls without any human intervention, according to its boss Katya Lainé, who works at VivaTech.

For her as for Victor Riparbelli, the use of actors’ voices without their consent is a red line.

In the case of Scarlett Johansson, “if they imitated her voice without her knowing, then it’s really bad,” responded Mr. Riparbelli.

But “if they don’t have the right to use (the voice) of someone who looks a lot like him, that sets a very strange precedent,” he said.

-

-

NEXT Valady. Jean Couet-Guichot and Gaya Wisniewski, two artists in residence within the region