As a general rule, it is easy to discriminate within a small set of words, but error rates naturally increase as the vocabulary size grows. The Japanese Large Vocabulary Continuous Speech Recognition (LVCSR) system (three-year project, 1997-2000), so-called JULIUS, funded by the IPA (Information-technology Promotion Agency) in Japan, shows the word accuracy of 94.4% and 93.1% in a word 3-gram and triphonemodel for vocabulary sizes 5000 and 20000, respectively[10].
