Efficient word-graph parsing and search with a stochastic context-free grammar

Word graphs provide a compact representation for the output of the acoustic decoder in a speech recogniser. This paper proposes an efficient algorithm for parsing the graph structure using a stochastic context-free grammar (SCFG). By parsing an entire graph in a single pass significant savings can be made over techniques that parse individual sentences, such as the common N-best strategy, or methods that parse paths through a graph. A backward Viterbi search is used to recover the parsed sentences from the graph. The full graph parsing algorithm is shown to be better than a heuristic search that parses only portions of the graph. On the resource management task a reduction in computation of 5 times over N-best is demonstrated. The integration of SCFG gives a reduction in recogniser word error rate of 9.8%.