TOKYO -- Japan hopes to see foreign visitors navigating the capital seamlessly during the 2020 Olympics with the help of small translation devices shaped like a nametag or earphone.
To make that vision a reality, a tech institute under the communications ministry will open up its artificial intelligence-driven translation tools for use by companies to make prototype products.
The National Institute of Information and Communications Technology, or NICT, has honed its tourism-oriented translation tools since June using deep-learning techniques. The tech now boasts roughly 90% accuracy in fields such as disaster prevention and shopping, and it often offers phrases more natural than those generated by Google Translate.
NICT's VoiceTra app can translate text to or from 31 languages, as well as listen and speak in 16, including English and Chinese. Its simultaneous interpretation function detects when a sentence ends and can start delivering a translation moments afterward. The institute hopes to use image recognition for interpreting speakers' facial expressions as well.
The government looks to set up AI servers as early as fiscal 2018 equipped with NICT's translation tech to receive voice data from companies' prototypes, waiving licensing fees for the tools at that stage of development.
The communications ministry aims to eliminate upfront investment costs -- which can reach millions of yen (1 million yen equals $9,355) for servers and licensing -- as well as wait times for licensing deals, thereby lowering hurdles to development. The ministry hopes to accelerate product creation, a process that has yielded only gradual results such as Fujitsu's target of fiscal 2018 to launch a clip-on translator device for medical professionals.
Access to NICT's core technologies could help businesses develop a range of products tailored to distinct designs, target languages or the ability to serve different numbers of people at once. For instance, many Japanese izakaya pubs and restaurants feature touchscreens for ordering at seats and tables. Interpretation technology could be incorporated into such devices.
Opening the institute's translation tools to the private sector also should make them more accurate. As the tools train with a wide range of data, including technical terms from various businesses, they can handle more fields and broaden their vocabulary.
U.S. tech companies including Google and Microsoft are striving toward multilingual translation tools trained through deep learning. Such businesses tend to guard their technology jealously. Japan, which has trailed in gathering relevant data, hopes to catch up in the development race by including private businesses.
NICT has teamed with around 180 domestic research institutions and businesses, including Fujitsu and Hitachi, on improving technology and pursuing practical applications. The institute launched a "translation bank" in September for collecting translation data from businesses and other bodies.