Google: AI language tools reach 300 languages, 7 billion people

Google frames this milestone as the product of two decades of work that began with Google Translate in 2006. At launch, the goal was simply to break down language barriers, and AI advances have since expanded Translate from a handful of languages to more than 250. But the company argues that translating text is not enough, because speech recognition systems historically ran a rigid pipeline — transcribing audio, processing the text, then synthesizing it back — that discarded tone, pacing, emotion and context. People laugh, hesitate and mix languages mid-sentence, as in Spanglish or Hinglish.

The response has been to train models such as Gemini on audio directly, and to treat language data as a community matter rather than a web-scraping exercise. Google says the web disproportionately represents a few dominant languages, so it turned to local partnerships: WAXAL with Makerere University and Digital Umuganda covers 27 Sub-Saharan African languages; Project Vaani with the Indian Institute of Science and Bhashini has collected more than 30,000 hours of speech across 109 languages; and the Amplify Initiative worked with more than 1,600 local experts. Open tooling such as Language Explorer, which continuously maps more than 7,000 languages, supports the same aim.

The stakes appear to hinge on reach rather than model quality. For more than 3 billion people, reliable internet access is still out of reach, so the on-device TranslateGemma models and feature-phone efforts like Viamo's AVA in Rwanda are the test of whether these capabilities actually arrive. Similarly, the accessibility work — including sign-to-text dictation on Pixel 11 — and the Māori place-name pronunciation work with local experts suggest Google sees cultural accuracy as part of the product, not a footnote. Whether the 1,000-language goal moves at the pace of the research or the pace of local data collection remains to be seen.