This paper analyses the incorporation of rule-based language models based on probabilistic context-free grammars into speech recognition systems. The important issue of efficiency is addressed and the parsing of word lattices is shown to be both effective and fast. It is shown that parse failure can occur when the lattice is incomplete. The application of a phrase level integrated grammar/bigram language model is investigated with particular relevance to handling these parse failures.
Bibliographic reference. Jones, G. J. F. / Lloyd-Thomas, H. / Wright, J. H. (1995): "Lattice parsing and application of integrated language models for speech recognition", In EUROSPEECH-1995, 1789-1792.