HTML5 Webook

19/64

NICT REPORT 17dations needed for the development of multi-lingual translation services), and made a significant contribution to accelerating the implementation of speech translation tech-nology in society.In cooperation with the Ministry of Internal Affairs and Communications, we have also introduced a “Hon’yaku (Translation) Bank” scheme whereby NICT can collect transla-tion corpora from diverse sources scattered throughout Japan. This scheme uses various means (including uploads via the Web) to ef-ficiently collect multilingual translations in multiple fields from around the country, and is expected to achieve greater precision in gen-eral-purpose machine translation (Fig.3). This Hon’yaku Bank scheme can be described as a new donation-based collection method.R&D of a simultaneous language interpretation platformWe have developed a prototype one-shot speech translation system as the first stage of our research and development aimed for completion in 2020. In its current system, Voi-ceTra must send a series of four requests to the server in order to process one transla-tion—one to perform speech recognition, an-other to perform text translation, one more to perform reverse translation, and a final re-quest to perform speech synthesis. With a one-shot speech translation system, the input speech is sent to the server as a speech translation request, the server performs all the necessary processing during this one single request, and then sends back the speech recognition, translation, reverse translation, and speech synthesis results in sequence as a single response. This results in a faster re-sponse speed than the current system where four round trips of requests and responses are required. In a comparative evaluation we performed in Europe, the current method had a reaction speed of about 6 seconds, while the one-shot speech translation method had a reaction speed of about 2 seconds. In Ja-pan, the reaction speed increased from about 2 seconds to about 1.5 seconds. In the future, we will incorporate the one-shot speech translation method into VoiceTra. We also plan to expand this system to work with con-tinuous speech input, which would enable the development of a system that can per-form simultaneous interpretation of lectures and the like, and to provide a research plat-form for simultaneous interpretation.Fig.1 : The outlines of speech synthesis methodsFig.2 : Results of perceptual evaluation of synthetic speech based on MOS (Mean Opinion Score)Fig.3 : The “Hon’yaku Bank” conceptThe NICT Team formed by Dr. Komei Sugiura, Dr. Aly Magassouba, and others from the Ad-vanced Speech Technology Laboratory of ASTREC, won the "1st Place (METI*1 Minister’s Award)" and the "JSAI*2 Award" in the Partner Robot Challenge Virtual Space at the World Robot Summit (WRS) 2018. WRS 2018 was held from Oct. 17 - 21 at Tokyo Big Sight and was sponsored by METI and NEDO*3 . In this challenge, seven domestic/international teams competed with each other and they were judged based on the achievement rates of the following 3 tasks: 1) multimodal lan-guage understanding task—how accurate can the service robot in virtual space understand us-ers' commands using non-linguistic information such as images; 2) gesture recognition task in the same virtual environment; and 3) multimodal lan-guage generation task. The NICT Team marked the best achievement rates in all 3 tasks leading themselves to their victory.* 1 Ministry of Economy, Trade and Industry *2 The Japanese Society for Artificial Intelligence*3 New Energy and Industrial Technology Develop-ment OrganizationThe NICT Team being awarded the “METI Minister’s Award" by Parliamentary Vice-Minister of Economy, Trade and Industry, Mr. Akimasa IshikawaNICT team won awards at the World Robot Summit 2018Research and Development