Of the roughly 7,000 languages spoken on Earth, linguists estimate that nearly half could disappear within a generation. The tools that record and revive them have, until recently, been painstakingly manual: dictionaries compiled by hand, grammar sketches built one elder at a time. A new open-source translation model, trained to handle around 200 languages the industry has long ignored, suggests that pipeline is about to get a lot faster.
The problem with "low-resource" languages
Machine translation has become routine for languages like English, Spanish, and Mandarin, because those languages sit on mountains of paired text: the same sentence in two languages, repeated millions of times. Endangered languages do not. A language spoken by a few thousand people, mostly in one valley, may have almost nothing written down at all, let alone neatly aligned with a major world language.
Researchers call these "low-resource" languages, and for years they were effectively off-limits to AI. The breakthrough was less a single algorithm than a shift in method: training on many languages at once, so that structural patterns learned from a well-documented cousin can be borrowed by a poorer one. A model that understands dozens of Polynesian tongues can make a respectable guess at a related one it has barely seen. This multilingual approach mirrors the way edge computing distributes intelligence: instead of one massive central model, knowledge spreads across a connected network of related tasks.
Why AI changed the equation
For communities racing to preserve a language before its last fluent speakers are gone, even an imperfect translator is useful. It can turn a scattered archive of recordings into searchable text. It can draft a first-pass dictionary that a human speaker then corrects in an afternoon rather than a year. It can let a child read a story in a language their grandparents spoke but their parents never fully learned.
"The model is not fluent. But it is fluent enough to do the boring part, and the boring part is what was killing these projects."
That last point matters more than it sounds. Language revitalization is less often defeated by a lack of passion than by a lack of time. Elders die; volunteers burn out; grant cycles end. Anything that compresses the grunt work extends the window in which a community can act.
Where it still falls short
The technology has real limits. A model trained on limited data can hallucinate, inventing words that sound plausible but mean nothing. It struggles with ceremony and song, where meaning lives in cadence as much as vocabulary. And there is a harder question underneath all of it: a language is not just a code to be decoded. It is a way of seeing the world, and no model can preserve that on its own.
The people behind these projects are clear-eyed about this. They describe the AI as a tool, not a savior. Used well, it buys time and reach. Used badly, it produces fluent-sounding nonsense that erodes trust. The difference, as ever, is whether the communities whose languages are at stake get to decide how the tool is used. On that front, the open-source approach, which lets communities run and audit the models themselves, is the most promising sign in a field long dominated by outsiders.
