Google says 300 languages, on-device translation, and a voice AI that runs over a phone call
The 86% figure is the headline. The more interesting engineering is aimed at the other billion.
Key points
- Google's language products now cover more than 300 languages spoken by 7 billion people, or 86% of humanity.
- Google Translate has grown from a handful of launch languages in 2006 to more than 250 today.
- A lightweight model called TranslateGemma runs across 55 languages directly on a phone, no internet needed.
- A voice assistant built with non-profit Viamo, powered by Gemini, has answered more than 2 million questions on basic feature phones in Rwanda.
- Sign Language-to-Text, starting with American Sign Language, is being built into Gboard and Live Transcribe on the Pixel 11.
Google has laid out how its language tools work today, and the numbers are large: products covering more than 300 languages, reaching about 7 billion people. That is 86% of the planet. The gap it wants to close is the other billion or so, most living where reliable internet is absent or intermittent.
My read, after years of watching translation demos: the flashy live-interpreter clip is not the story. The story is a translation model small enough to run on a cheap Android with no signal, and a voice assistant that works over a plain phone call. That is where language tech stops being a party trick and starts being infrastructure.
What did Google actually announce?
A bundle of language projects, from real-time speech translation to tools for sign language and feature phones. Google Translate is still the headline product, now at more than 250 languages after launching in 2006.
The newer pieces sit on top of Gemini. Gemini 3.5 Live Translate handles spoken translation across 70 languages and more than 2,000 language pairs, built to cope with code-switching, the habit speakers have of blending two languages mid-sentence as Spanglish or Hinglish speakers do. Gemini 3.5 Transcribe converts raw audio into formatted text and powers Rambler, a Gboard feature that strips filler words and tidies grammar as you dictate. We covered Google's expanding Gemini voice work when the company added voice control to Gmail, Docs, and Keep on 3 September.
How does it work without the internet?
Google shrank the model. TranslateGemma is a family of small open translation models built from Gemini and trained across 55 languages, designed to run on the handset rather than in the cloud.
That matters because Google notes that more than 3 billion people still lack reliable internet. A translator that needs a signal is useless in a rural clinic on 2G. One that lives on the device is not. On-device AI has been a consistent thread in our coverage since 13 July, and this is one of the more concrete deployments we've seen.
For people without a smartphone at all, Google is backing Viamo, a non-profit whose Ask Viamo Anything service lets someone on a basic feature phone dial in and talk to a Gemini-powered assistant over a normal voice call. A pilot in Rwanda has handled more than 2 million questions.
Where is the data coming from?
From local partnerships, because the open web mostly speaks English. Google's Universal Speech Model was trained on 12 million hours of audio and uses cross-lingual transfer learning, a technique that applies patterns from data-rich languages to languages with far less recorded material.
| Project | Partners | Scale |
|---|---|---|
| WAXAL | Makerere University, Digital Umuganda | 27 Sub-Saharan African languages, 100 million+ speakers |
| Project Vaani | Indian Institute of Science, Bhashini | 30,000+ hours across 109 Indian languages, 155,000+ speakers |
| Amplify Initiative | 20 universities on 4 continents | 15,000 multimodal data points from 1,600+ local experts |
What about people who don't use speech?
Google is building Sign Language-to-Text, trained across more than 50 sign languages. It will power sign-to-text dictation in Gboard and Live Transcribe on the Pixel 11, starting with American Sign Language to English. Roughly 70 million people worldwide use sign language as their primary means of communicating. Making a keyboard that accepts signed input, rather than forcing a signer to type, is a genuine shift in who these devices are designed for. Two weeks ago we reported on DiscoSign, Apple's attempt to handle the conversational elements that sentence-by-sentence sign language systems routinely drop; Google's approach is narrower in scope but ships on a mass-market device.
One telling detail: in New Zealand, Google worked with Māori language experts to correct how Google Maps pronounces place names. A navigation app that mangles a town's name isn't neutral. Getting it right is modest engineering and real respect.
