No, you still can’t talk to dolphins or aliens. But with Google Translate, you can now talk in your native tongue across most corners of the globe.
Or at least read signs.
In January, Google Translate introduced the capability for instantaneous text translation in seven languages (English, Russian, Spanish and four European languages). Snap a pic of a street sign in Moscow, and see it appear in English. The feature was the fruit of Google’s 2014 acquisition of Word Lens, which made a translation app for Google Glass.
On Wednesday, the feature tacked on an additional 20 languages. Four of them — Indonesian, Filipino, Hindi and Thai* — are spoken in some of its largest potential markets, evidence of the search giant’s ongoing courtship of Internet newcomers beyond the lingua franca. The new feature also works when a phone lacks a Web connection.
With the new feature, Google Translate marks another first: The use of artificial neural networks. As Re/code reported earlier, advanced algorithms that push machines to train themselves, built inside Google’s deep learning unit, bleed into over 100 teams inside the company. For this one, Translate machines learn to take images out in the wild, then pick out what’s a letter and what’s not. (Interestingly, here the Translate team used convolutional neural nets, an image recognition technique closely associated with Yann LeCun, Facebook’s AI chief.)
Google scored a ton of lavish press with its “Inception” neural nets — those far-out image simulators you see all over the Web. Otavio Good, a wonderfully-named Translate engineer, brought them up in his explanation of how neural nets work here: “Yes, they’re good for more than just trippy art — if you’re translating a foreign menu or sign with the latest version of Google’s Translate app, you’re now using a deep neural net. And the amazing part is it can all work on your phone, without an Internet connection.”
Google is not the only tech giant trying to crack instant translation. Microsoft’s Skype also has a real-time language translation product, which works for fifty languages.
* For 24 of the languages, Google can do two-way translation — from English to that language and back. But it’s only one way with Hindi and Thai, two non-Latin-based forms of writing. Which reinforces the truism that machine learning is hard.
This article originally appeared on Recode.net.
