Stochastic language models for Chinese speech recognition based on Chinese spelling

The rate of speech recognition can hardly be improved when it is as high as 90%, unless the speech understanding technique is used. In this paper, a new approach to Chinese speech understanding (a spelling based stochastic language model approach) is proposed and has been used to solve the problem of unrestricted speech understanding, which the classical method (rule based approach) can not. It can be used to eliminate two thirds of all the syllable errors while reducing the processing time tremendously. As a result, a Chinese speech recognition system has become commercially available