In this paper, a new natural language processing approach based on the statistics of corpus is proposed and has been successfully applied to THED-919 Chinese speech recognition system to eliminate acoustic recognition errors and to translate spellings into Chinese words. The accuracy rate of spelling-word translation of unrestricted text is 98.4% and 2/3 of acoustic recognition errors are eliminated. Keywords: Speech Recognition, Speech Understanding, Markov Chain, N-Gram Model.