After the image (particularly with Dall-E) and the text (with ChatGPT), the voice seems to be the new playing field of artificial intelligence. Microsoft presents VALL-E, a tool capable of reproducing everyone’s voice from a sample of just three seconds. The promise of the software is to be as faithful as possible in its imitation.
To do this, Microsoft has fed its artificial intelligence with 60,000 hours of data spoken in English, explains the American site Ars Technica. The great strength of VALL-E is being able to transcribe a person’s tone and emotion. Thus it is possible to obtain an incorporated reading even if the spoken words do not appear in the original sample. The American company has published examples on a dedicated web page.
Dangerous uses?
Of course, the generated voice will be all the more realistic the longer the initial sample is. Audio files of three seconds are the limit beyond which a mimicry can occur. But more accurate results can be obtained by delivering more material to VALL-E.
Like all content generated by artificial intelligence, this technology opens the way to usurpation. Political figures or celebrities could see non-consensual messages (called deep fakes) expressed from a sample of their voice.
VALL-E also raises serious security concerns. As the Windows Central site specifies, certain services (such as banks) use the voice of their users as their password.
Finally, it is especially artistic activities that could suffer the most. From a single sample, VALL-E would be capable of managing tasks currently reserved for humans. In particular dubbing of movies or series, or even audiobooks.
For now, Microsoft does not offer Internet users to generate their own speech synthesis. The company assures that it will also develop a tool in charge of detecting a “false voice”, in order to limit abuse as much as possible.
Source: BFM TV
