Google is developing an AI that will speak (almost) every language in the world

Targeting support for the thousand (!) most spoken languages, Google wants its AI to be able to speak – and respond – to most people on our planet. If the technology works, it will make its services near universal. And will erase, perhaps, the domination of English.

After building the Tower of Babel, men would have suffered the wrath of God, who would have divided us by introducing languages. Google seems to want to fix this mythological injustice with a new artificial intelligence that would speak “ the thousand most spoken languages ​​on Earth “. Just that – and take it in your teeth, God.

Before this tower was built, we were all supposed to speak the same language.
“The Tower of Babel”, Pieter Brueghel the Elder, 16th century.

Google is already a world champion in translation with its Google Translate service, a technological brick that it integrates everywhere, but the work is far from perfect. Many languages ​​are poorly or not at all supported, translations are often imperfect. And for many services, English dominates. It took several years for Google to integrate other languages ​​into the exclusive “Recorder” text transcriber of its Pixels.

Babel may have been destroyed in a day, but Rome took longer to build. As a human company, Google therefore began by setting an initial objective of taking into account 400 languages ​​by its new AI. When we know that there are 22 official languages ​​in India, we easily understand why these 400 languages ​​are only a stage. Africa alone would have around 2,000 languages ​​and dialects (but the numbers go up to 3,000!), which should keep engineers busy for a while. Google is also not the only company on the spot, but the most advanced project, that of Meta, is satisfied with “only” 200 languages.

Read also: Meta has developed a translation AI capable of translating the most exotic languages ​​(July 2022)

Google’s AI will not combine several bricks combining languages, but will consist of a single model that will be trained to know them all. ” By having a single model exposed and trained in several languages, we benefit from better performance, especially on languages ​​with few resources Google AI VP of Research Zoubin Ghahramani told The Verge. Who explained that the AI ​​they are developing will rely on the similarities between certain languages ​​to strengthen its robustness.

Because Google’s AI is not only designed to allow a Tamil merchant to sell tea to an Uzbek trader. Google sees its tool as a super language capable of “transforming commands (sent to a robot, editor’s note) into code or solving mathematical problems”. All this in addition to the translation. Making AI a digital model capable of solving many problems. For different services: even if Google has not communicated targeted products for its AI, all the bricks of the American giant could benefit from it.

For the time being in the state of applied research, the Google project does not yet have a defined public roadmap. But besides the daunting task of creating a super multilingual AI, a new task if ever there was one, other known challenges await. Like avoiding creating a monster, as is often the case when training this kind of pattern. Because, product of men, AIs are sometimes in our image, that is to say with a bad background. A touch of divine intervention would no doubt be welcome!

Source :

The Verge

Leave a Comment