Meta announced a new AI model called Voicebox yesterday, one it says is the most versatile yet for speech generation, but it’s not releasing it yet: The model is still only a research project, but Meta says can generate speech in six languages from samples as short as two seconds and could be used f...
Meta announced a new AI model called Voicebox yesterday, one it says is the most versatile yet for speech generation, but it’s not releasing it yet: The model is still only a research project, but Meta says can generate speech in six languages from samples as short as two seconds and could be used for “natural, authentic” translation in the future, among other things.
Bloody hell it's amazing how desperate this marketing attempt is. I've got an AI that'll blow your faces off with its output but I can't show you because, well, your face will resemble Gus Fring after several weeks in an acid bath. I can show you if you pay me a Huffmanian sum for API access though, but only if you sign an NDA and promise not to say mean things about us.
Ah, but allow private access for wannabe authoritarians who will use it to create fake ad campaigns.
No public access = no scrutiny from researchers nor watchdogs.
One thing I am hoping for with "AI" tech is to have better language teaching software. It would be crazy to have an AI teacher correct you mid conversation or be able to adapt to what you are struggling with.
Not really AI but if you want to learn lojban you get most of the way: The language actually has a formal grammar so it's possible to write bog-standard software that doesn't care when you say "You must have patience, my young Padawan" instead of "Patience you must have, my young Padawan". It's going to tell you which it expected but not beat you over the head with it.
This is actually something I was working on at my last job: Using speech recognition to capture oral reading errors and give feedback to children. It's not a large field but definitely one of the more noble AI applications. One of the major downsides is that because of privacy concerns you will want to run models locally instead of simply accessing 3rd party (OpenAI/Meta) models through an API, which introduces performance limits.
Translation... "This 'AI' is hot garbage that would make us look bad if we released it in its current state, but we want to hang out with the cool AI kids... So instead, we'll say that it's SO advanced and so good that it's DANGEROUS for the public. That way we look cool, ethical, and mysterious and definitely NOT that we blew our wad on the metaverse last year and are reaching for anything to climb out of the hole we dug..."
“Our technology is soooo jaw droppingly powerful, we must warn the public!” It just seems a little self serving. ‘Critihype’ (Motherboard?) was a term I heard recently that sums this up nicely.
You can already make deepfake videos from a single target face image, and now you can make AI generated voice from two seconds of sample? The future is going to be interesting to say the least.
If they actually had it, I'm pretty sure they would release a demonstration video at least - that doesn't require sharing the data or the code, so I can't imagine why they wouldn't.