The high cost of developing core technology in speech recognition highlights the need for freely available state-of-the-art software. We have released an initial version of an LVCSR decoder that supports cross-word context-dependent phone modeling and lattice rescoring. The core search engine uses a variation of the Viterbi algorithm to efficiently manage word, phone and state-level hypotheses. The decoder employs lexical trees to handle multiple pronunciations of words, and also supports general network decoding. Preliminary evaluations on the WS'97 dev test partition of SWITCHBOARD (SWB) yielded a 45.2% WER. The decoder is also shown to be competitive in computational requirements.