Newsletter
Google Translate has had the ability to use a neural network to improve translations for around two years, and now the capability is available offline, thanks to on-device machine learning. The change is effective immediately, and will work with any downloaded language packs in both iOS and Android. It is worth noting that you’ll need a device that has onboard machine learning in some capacity in order to take advantage of the new feature, which means just about any flagship device from 2017 or newer. Qualcomm’s Snapdragon 835 mobile processor was the first with the capability, and the likes of Samsung, MediaTek and Huawei followed shortly after.
The way it all works is by using neural networking to examine and translate entire sentences in context. This helps to avoid running afoul of grammatical conventions, and gives some clues for words that may be similar on paper and have different meanings in different contexts. The result is translations that are easier to read and sound more natural. The language packs that make these offline translations usually run under 50 megabytes each, which means that a globetrotter could potentially keep the files for a range of destinations around the world on hand without having to sacrifice anything else stored on their device of choice. Offline translation is most useful for situations where you can’t get a data connection or you can only get a limited or low-speed one.
Onboard machine learning in mobile devices has promised to bring a ton of unique advancements to consumer devices like smartphones and tablets, and this is one of the biggest use cases thus far to prove that point. It is worth noting that mobile machine learning and neural networking existed before the advent of on-device machine learning capabilities, but those functions were limited to devices that were connected to the internet, and required either a strong server on the other side of the connection or a large number of other users running the same app or service to act as nodes. One of the most notable examples of such is the Nigel AGI project being worked on by Kimera Systems.
