Google’s online translations have been powered by neural machine translation (NMT) since 2016, and today the company is rolling out its neural net-driven approach to more accurate, natural-sounding translations for Google Translate iOS and Android app users to carry out translations offline in 59 languages.
Offline NMT was made by the Translate team in conjunction with the Google Brain team using TensorFlow, Google product manager Julie Cattiau told VentureBeat in a phone interview. Unlike for other Google apps, 95 percent of Google Translate’s user base is outside the United States, in countries like India, Brazil, and Indonesia, Cattiau said.
“So we hear a lot from our users that it’s great to have high quality online, but a lot of them are either unable to access an internet connection or they would prefer to save on their data plan. So we made it a priority over the past year and a half to basically squeeze our NMT models onto people’s devices,” she said.
Rather than the previous machine learning approach that provided interpretation by scanning phrases of a sentence, offline translations with NMT analyze entire chunks of text at once, allowing for more natural-sounding, grammatically sound, context-aware translations.
When connected to the internet, Conversations mode in the Google Translate app can provide on-the-spot voice translation. However, NMT offline translation launches today with text-only translation; it does not extend to features in the Translate app that allow you to interpret the menu you take a picture of or translate people’s voices. Conversation mode in Pixel Buds, Google’s first pair of earbuds, drew comparisons last fall to the Babel fish in Hitchhiker’s Guide to the Galaxy.
In order to make real-time voice translation possible offline, Google will have to make other elements of AI that combine to enable Conversations mode available offline too, such as speech recognition and synthesis of words from text back into speech.
“Each of those parts needs to be built on-device for the full experience to work, and that’s definitely something we want to launch,” she said. “There’s no date announced at the moment, but text translation is definitely one of the building blocks that will lead towards having an end-to-end offline translation for speech.”
No app update is necessary to get offline neural machine translations. Google Translate users who previously downloaded offline translation packages will see a banner encouraging them to click there for better translation, while Translate users new to offline translations will have to go into the app and select the languages they want to use offline. Each language package will take up roughly 35-45 MB, roughly equivalent in size to previous offline packages but higher quality.
“We can’t run these very consuming models that requires a lot of computing power on $50 phones, so the trick with this launch was for engineers to squeeze the models and make them run on very low-end Android devices. That was the challenge of this launch,” Cattiau said.
User should notice a difference in quality from previous offline translations, but online translations will still be more accurate than offline translations, Cattiau said, as concessions were made to reduce language packages in size.
Like the previously used phrase-based machine learning approach, NMT leverages hundreds of millions of example translations of things like articles, books, documents, and search results.
Google’s language prowess doesn’t just improve its Translate app. The Alphabet subsidiary has committed to making Google Assistant available in more than 30 languages by the end of the year, a number that far surpasses Alexa’s four languages and Siri’s ability to speak 20 languages.