Understand any language, instantly, with AI

Clovis McEvoy
March 27, 2024
Meta's No Language Left Behind AI model can instantly translate 200 languages with state of the art quality.

Revolutionary progress in AI translation is breaking down barriers, improving access, and preserving the world's cultural heritage.

Humanity is on the cusp of achieving one of its oldest goals: universal translation.

The struggle to understand one another has always been with us, and from the Rosetta Stone to the earliest digital translators, each era has harnessed the technology of the day to solve this fundamental problem. Now, the AI revolution is poised to break the language barrier once and for all to deliver a true milestone for our species.

The last few years have brought stunning achievements. Sign language translation apps are opening up new levels of accessibility, Google is creating a tool that not only redubs video into a different language, but synchronises the speaker’s lips to match the new translation, whilst the Samsung Galaxy S24’s Live Translation feature offers two-way translation of voice and video calls, text messages, and in person conversations – all processed locally on the phone without the need for an internet connection.

But these recent achievements only hint at what’s coming next.

A replica of the Rosetta Stone displayed as part of the Treasures of the World's Cultures exhibition at Centro Exposiciones Arte Canal in Madrid, Spain.

There are some 7000 languages currently spoken around the globe – yet the most ubiquitous translation app, Google Translate, can process a mere 133. This gap in coverage is what researchers refer to as the ‘last mile’ of machine translation, and it’s a challenge they have been stuck on for some time.

Stopping them is a data problem. Historically, training a translation model has required copious amounts of written and recorded speech. Common languages like English and French were quickly mastered thanks to the vast trove of literature, film, and radio that researchers could draw from. However, ‘low-resource’ languages – languages with far fewer published materials – have proven much harder to crack.

Fula, with 37 million speakers in West Africa, and Kirundi, the native language of Burundi with 14 million speakers, are some of the biggest missing languages on Google Translate. Cantonese and Shanghainese, with over 100 million cumulative speakers, are also missing, though it’s worth noting that Google is blocked by the Chinese government, where most of those speakers live.

It's an issue with serious consequences: accessible translation is essential for disaster and aid relief efforts across the globe. For migrants and asylum seekers, the accuracy of a translation app can mean the difference between sanctuary or deportation.

The new generation of large language models represent a breakthrough for translation. According to Google’s Chief Scientist, Jeff Dean, the company’s Gemini 1.5 Pro model was able to translate Kalamang – a language thought to have fewer than 200 active speakers – using only a single grammar manual as a data source.

Beyond Google, Meta’s No Language Left Behind project has quickly hit the milestone of 200 translated languages (surpassing Google’s 133), including 150 low-resource languages, and have made their revolutionary translation models open source and public for anyone to use for free.

Meta's No Language Left Behind model, NLLB-200, can instantly translate 200 languages with state of the art quality. It's already improving translations on Facebook, Instagram, and Wikipedia.

Such progress is not only important for communication. This inclusion is vital for language preservation. UNESCO estimates that over 3000 languages could die out before the end of this century. This diversity is a core part of human expression; each language represents a history of different life philosophies, concepts, and cultures. To lose it would be an immense loss to human civilisation, locking our society into increasingly narrow perspectives on life.

Preventing this means that people, especially young people, need to be able to use their preferred or native language in day to day life and to access digital resources on their own terms. Whilst 99% of the online population is able to translate the internet into a language they know, millions of people do not access it in their mother tongue. The only real option is to choose from a relative handful of dominant languages; a majority of websites are in English.

Here, too, technology is making a pivotal difference. A growing number of indigenous communities are taking it upon themselves to preserve and revitalise their language using the power of AI, and Meta’s open source models are currently being used to translate Wikipedia into a number of endangered languages.

AI models represent a breakthrough for translation.

Meta’s ultimate goal is to create a single AI model capable of true, universal language translation. For the first time in history, the prize is clearly within reach.

It is hard to overstate what a monumental achievement it would be.

Communication carries within it all the potentials for collaboration, compromise, fraternity, and love. Translating language is an essential act of cultural sharing, prized by generations of philosophers, inventors, and scientists.

Where ancient Sumerians once chiselled translations of the Epic of Gilgamesh into stone tablets, seamless translation across every language and between all the peoples of earth is now no longer science fiction but an immanent reality.

The world has 7,000 languages, but most people only have access to a few hundred of them.
Click to view our article about art.
Written by
Clovis McEvoy
Click to view our article about art, music, film, and storytelling..
More about
Click to view our article about art.
More about

Clovis is a New Zealand born writer, journalist, and educator working at the meeting point between music and technological innovation. He is also an active composer and sound artist, and his virtual reality and live-electronic works have been shown in over fifteen countries around the world.