In this paper we describe an algorithm for generating word networks in a continuous speech recognition system. Recently, iV-best search strategies have gained considerable popularity and have been used for multi-stage searches including interfacing speech recognition and natural language systems as well as applying more computationally expensive constraints in later stages. However, examination of TV-best lists reveals significant overlap between different hypotheses, with differences typically localized to regions where the acoustic signal is not robust. In order to improve both computational and representational efficiencies, we have developed a word network search. This search is very similar to our A* iV-best search, but contains an additional path-merging step. The resulting word networks contain the same N complete hypotheses that are within a specified score threshold of the best complete score, but in a much smaller form that is faster to compute. These word networks can then be used as search spaces by subsequent search stages.
Bibliographic reference. Hetherington, I. Lee / Phillips, Michael S. / Glass, James R. / Zue, Victor W. (1993): "A* word network search for continuous speech recognition", In EUROSPEECH'93, 1533-1536.