Development of Indonesian Large Vocabulary Continuous Speech Recognition System within A-STAR Project

The paper outlines the development of a large vocabulary continuous speech recog-nition (LVCSR) system for the Indonesian language within the Asian speech transla-tion (A-STAR) project. An overview of the A-STAR project and Indonesian language characteristics will be briefly described. We then focus on a discussion of the develop-ment of Indonesian LVCSR, including data resources issues, acoustic modeling, lan-guage modeling, the lexicon, and accuracy of recognition. There are three types of In-donesian data resources: daily news, tele-phone application, and BTEC tasks, which are used in this project. They are available in both text and speech forms. The Indonesian speech recognition engine was trained using the clean speech of both daily news and tele-phone application tasks. The optimum per-formance achieved on the BTEC task was 92.47 % word accuracy