This paper presents a fast and accurate approach to continuous speech recognition for Chinese language with very large vocabulary. Since Chinese language is not alphabetic and the input of Chinese characters into computers through keyboards is difficult, such speech input techniques for dictation purposes are highly desired. Considering the special characteristics of Chinese language, the approach proposed in this paper is based on a two-stage recognition concept. The first fast matching stage utilizes the monosyllabic structure of Chinese language and provides a word lattice to constrain the search space of the subsequent stage, while the second detailed matching stage successfully integrates the acoustic word models and the Chinese language model. This approach is therefore much more accurate and faster than our previous version for such a task[l]. The current experimental system runs in real-time on SPARC-10 workstation with Chinese character accuracy on the order of 90.2%, while the previous version takes more than twice of the time with accuracy of 88.7%.
Bibliographic reference. Ho, Tai-hsuan / Wang, Hsin-min / Chien, Lee-feng / Chen, Keh-Jiann / Lee, Lin-shan (1995): "Fast and accurate continuous speech recognition for Chinese language with very large vocabulary", In EUROSPEECH-1995, 211-214.