Meta Develops Groundbreaking AI That Can Replicate Human Voices But Won't Release It

Meta has created an AI called Voicebox that can generate audio that sounds like real human speech in six different languages.
Voicebox is able to take in text it has never seen before and create natural-sounding speech with proper inflection and context, just as a human reader would.
Meta believes Voicebox could help visually impaired people and those with speech impairments but will not release the technology publicly due to concerns about how it might be misused to generate deepfakes.
Voicebox represents a huge leap forward for generative AI and has implications for audio editing, translation, and sampling in addition to accessibility.
Despite its promise, Meta thinks the risks of releasing Voicebox currently outweigh the benefits and is calling on other researchers to help solve issues around verifying and preventing synthesized audio misuse before releasing such a realistic generative audio model.