Like others, OpenAI is interested in synthetic voices, to allow an AI to imitate the voice of a person. Sam Altman’s group claims to have found a way to generate a voice clone of a human in just 15 seconds.
In the near future, will we be able to give ChatGPT the voice of our best friend, a parent or, why not, our own voice? The voice model created by OpenAI at the end of 2022, simply called “Voice Engine”, seems more and more effective. Giants like Apple can generate a voice in 15 minutes, OpenAI says it can do it in 15 seconds. All he would need to do is listen to any person for this amount of time to generate a synthetic voice capable of even imitating emotions.
Beyond the feat, an interest in accessibility
It is in a blog post published on March 29 that OpenAI details its voice engine. The company, which calls itself “cautious”, imagines several uses for its revolutionary model. Among them :
- Assistance in reading, to help people concentrate with a familiar voice.
- Real-time translation, to allow you to learn or speak another language with your own voice.
- Accessibility, targeting people who cannot or can no longer speak.
The few extracts published by OpenAI are impressive, since we cannot distinguish any differences between the original extract and the generated voice. Voice Engine is already the model that powers ChatGPT Voice, as well as translated podcasts in Spotify.
You won’t be able to clone a celebrity
“We recognize that producing speech that sounds like people’s voices poses serious risks, which are particularly significant in this election year. » said OpenAI at the end of the press release, to reassure about its intentions. Voice Engine is not available to the general public to avoid abuses and the companies that use it undertake not to usurp the identity of others.
To avoid negative uses, OpenAI added an audio watermark to its technology, which a machine can detect. “We believe that any large-scale deployment of synthetic voice technology should be accompanied by voice authentication experiments to verify that the original speaker knowingly adds their voice to the service and a list of prohibited voices to allow detect and prevent the creation of voices too similar to personalities » adds the company.
Do you want to know everything about the mobility of tomorrow, from electric cars to e-bikes? Subscribe now to our Watt Else newsletter!