Accuracy and speed are the main issues to consider when designing a large vocabulary speech recogniser. Recent experience with the Wall Street Journal (WSJ) corpus [5], has shown that high recognition accuracy requires the use of detailed acoustic models in conjunction with well-trained long span language models. In this paper we present a two-pass decoder architecture which extends an original [4] one-pass design. The initial pass consists of a time syn- chronous backward search in a pre-compiled network using simplified acoustic models and a null grammar. The forward pass can function as a stand-alone one-pass decoder capable of using cross-word context-dependent models and long span language models. The capabilities of this framework are empirically examined in terms of recognition accuracy vs speed on the Wall Street Journal database.