Did you know about deepfakes, those videos where an impersonator copied a celebrity’s face to make them say anything? Well, the startup ElevenLabs has just presented the first version of a program that will complete the paraphernalia: an artificial intelligence capable of faithfully reproducing a person’s voice.
Based on the Dall-E or ChatGPT model, the tool is very simple: just enter a text that will then be recited by the pre-recorded model, synthetic but realistic voices (only in English at the moment). The idea is, for example, to easily produce audiobooks.
But the real coup de force of this startup, founded by ex-Google and Palantir employees, is to be inspired by simple audio excerpts, in MP3 format, that the user can send to the algorithm. In all, it is possible for me to work on fifty excerpts for the same voice.
Here, for example, is what the editorial team at Tech&Co was able to create using the voice of Donald Trump, who suddenly became a fan of Eddy Mitchell.
And here is Barack Obama complaining about the Paris metro.
This tool, which however rests on a fragile legal basis, was obviously quickly taken up again in the forums where users multiplied more or less tendentious jokes. “As we see our technology being used massively in positive ways, we are also seeing an increasing number of voice cloning misuse cases,” ElevenLabs tweeted.
According to the US site Vice, clips in particular were broadcast containing the voices of celebrities, including British actress Emma Watson, uttering racist insults. For its part, ElevenLabs promises to think of a solution to better manage its tool. Among the ideas: impose an identity verification or a means of payment to be able to trace the person who generates the sound.
Source: BFM TV
