Meta's New AI Models Can Recognize and Produce Speech For More Than 1,000 Languages (technologyreview.com) 19
Meta has built AI models that can recognize and produce speech for more than 1,000 languages -- a tenfold increase on what's currently available. It's a significant step toward preserving languages that are at risk of disappearing, the company says. From a repprt: Meta is releasing its models to the public via the code hosting service GitHub. It claims that making them open source will help developers working in different languages to build new speech applications -- like messaging services that understand everyone, or virtual-reality systems that can be used in any language. There are around 7,000 languages in the world, but existing speech recognition models cover only about 100 of them comprehensively. This is because these kinds of models tend to require huge amounts of labeled training data, which is available for only a small number of languages, including English, Spanish, and Chinese. Meta researchers got around this problem by retraining an existing AI model developed by the company in 2020 that is able to learn speech patterns from audio without requiring large amounts of labeled data, such as transcripts.
Cool (Score:2)
But is it intelligible?
Pointless producing reams and reams on unintelligible> shit, now is it?
Re: (Score:2)
Still no edit after 25 years..... on - of
It's amazing (Score:3)
Are we sure? (Score:2)
Given Zuck's past shenanigans, it's not beyond the realm of possibility that he's hired a few thousand foreign-language speakers and is having them pretend to be an AI when someone starts a chat.
Re: (Score:2)
Even Zuckerberg isn't cruel enough to shove actual people through the pipes into your house and make them live in one of those little beige boxes.
Good god (Score:5, Funny)
They trained it on two new data sets: one that contains audio recordings of the New Testament Bible and its corresponding text taken from the internet in 1,107 languages, and another containing unlabeled New Testament audio recordings in 3,809 languages.
If you ever wanted to know how to say "Yea, verily" in a thousand languages, you're in luck.
Religion (Score:2)
Of course we have to train the data with a bible because it's not enough that we have to fight about religion in the real world, we have to create a bias within the "artificial" intelligence just how stupid it needs to be in order to be just like us.
Re:Religion (Score:4, Interesting)
My guess is that they chose the Bible because it has been translated into far more languages than any other book, and contains a lot of text.
1. The Bible (3350 Languages)
2. The Little Prince (380 Languages)
3. The Adventures of Pinocchio (260 Languages)
4. Tao Te Ching (250 Languages)
5. The Communist Manifesto (200 languages)
https://www.translateday.com/m... [translateday.com]
Re: (Score:3)
They missed a great trolling opportunity to use The Communist Manifesto.
Re: (Score:2)
Can you think of a better dataset? Did they translate Mission Earth into more than 3000 languages, full audio recordings in 1000 of those? Like it or not, this is a modern Rosetta Stone. We're extremely lucky to have this kind of resource.
I have no idea what "bias" you think this will introduce. Do you think that this will make your AI religious or something? Don't worry. That's complete nonsense.
Re: (Score:2)
You have a better idea? As ac22 points out, Christian missionaries are really the ones who have paid attention to small languages, now numbering in the thousands.
There are to be sure secular linguists who work on tiny languages, but they do not generally produce comparable corpora (i.e. corpora which are similar to those in other languages, which presumably gives the AI a leg up), nor do they work on as many languages as Bible translators have. Linguists who work on minority languages *tend* (with many ex
Re: (Score:2)
You shouldn't need specific examples of every possible word. These are models, after all, not datasets.
Mee too (Score:2)
Link to code (Score:4, Informative)
The article is missing a link to the actual models, code and article.
I found this that seems to be the relevant link: https://github.com/facebookres... [github.com]
Meta is still releasing the code (Score:2)
So hold your horses.
Re: (Score:2)
Nevertheless, the link I provided provides a link to the original scientific article, blog and other useful info that the OP article failed to mention.