This paper describes a method for the construction of a word graph (or lattice) for large vocabulary, continuous speech recognition. The advantage of a word graph is that a fairly good degree of decoupling between acoustic recognition at the 10-ms level and the final search at the word level using a complicated language model can be achieved. The word graph algorithm is obtained as an extension of the one-pass beam search strategy using predecessor dependent word models. The method has been tested successfully on the 20,000-word Wall Street Journal task (American English, continuous speech, 20,000 words, speaker independent). The word graph density could be reduced to an average number of about 10 word hypotheses, i.e. word arcs in the graph, per spoken word without loss of performance.