In this paper, we described a speech recognition algorithm using word trigram models that reduces the memory and computational cost, and reported the obtained experimental results. A recognition rate of 40% for text-open data and of 78% for text-closed data was obtained.