This paper describes the implementation of a large vocabulary speaker independent continuous speech recognition system for the DARPA RM task. The continuous HMM approach is used with mel-cepstra plus time differenced mel-cepstra and a new subword HMM topology. A set of within-word context dependent phone units are generated from a simple context merging scheme by counting and trained by both Baum-Welch and Segmental K-means algorithms. Only 6 continuous mixture diagonal Gaussian PDFs are used in each model. The recognition algorithm is based on a frame synchronous Viterbi beam search, which produces a lattice of words during forward searching, and N-best sentence candidates can be extracted in the reverse tracking, subject to the different levels of knowledge constraints. On the 991-word DARPA RM task, we present the preliminary results on the Feb-89 test set and the original SPHINX test set.