Google Translate's massive 110-language leap

If you’ve ever tried looking up Cantonese on Google Translate, you know the frustration. Despite having over 80 million speakers worldwide, it has been conspicuously missing for years.

That finally changes now. Google just added 110 new languages to Google Translate in what is the platform's single largest expansion to date. The update brings translation to more than 614 million speakers, representing roughly 8% of the global population.

Powering this massive rollout is Google’s PaLM 2 large language model.

Cantonese, reviving tongues, and African languages

Cantonese has long topped the wishlist of Translate users. Google noted that training models for it was tricky because written Cantonese often overlaps heavily with Mandarin, making clean training data hard to isolate.

Beyond Cantonese, the expansion covers an incredibly diverse linguistic spread. About a quarter of the new additions are African languages, including Fon, Kikongo, Luo, Ga, Swati, Venda, and Wolof, marking Google's biggest African language push ever.

Then there are the revival stories. Manx, the Celtic tongue of the Isle of Man, nearly died out when its last native speaker passed away in 1974. A local revival movement has since brought it back to thousands of speakers, and now it lives inside Google Translate.

Other notable adds include Afar, a tonal language spoken in Djibouti, Eritrea, and Ethiopia that saw the highest volunteer community contribution; NKo, a standardized West African script created in 1949; Tamazight (written in both Latin and Tifinagh scripts); and Tok Pisin, Papua New Guinea’s English-based creole.

How PaLM 2 cracks low-resource languages

Handling languages with almost no digital footprints is notoriously difficult for traditional machine translation. Google leaned on PaLM 2's cross-lingual capabilities to bridge the gap.

Basically, the model learns closely related languages together. If data is scarce for Awadhi or Marwadi, PaLM 2 leans on its knowledge of Hindi. It used a similar trick with French-based creoles like Seychellois and Mauritian Creole.

For languages without a unified standard, Google chose the most common varieties while blending regional flavors. For Romani, the system defaults toward Southern Vlax Romani, the most common variant online, while incorporating pieces of Northern Vlax and Balkan Romani.

Why it matters

For millions of people, this is real, everyday utility. Businesses dealing in emerging markets across Africa and South Asia, diaspora families trying to talk to grandparents, and travelers now have an instant bridge.

We think this move is a masterclass in where large language models actually deliver quiet, immense value. Traditional translation models starved to death without massive, hand-labeled parallel datasets, which endangered or low-resource languages never had. By using PaLM 2 to transfer knowledge across related dialects, Google solved a data bottleneck that stalled machine translation for decades.

It also brings Google closer to its 1,000 Languages Initiative. These languages won't drive immediate ad revenue, but they position Google as the essential digital infrastructure for the next billion users.

The real test will be accuracy. Standardizing diverse dialects into a single model output is bound to spark local arguments over "correct" grammar, and transfer learning can sometimes produce hallucinations that look deceptively convincing. For now, it is an incredible tool for connection, but maybe hold off on using it for legal paperwork.