First European Conference on Speech Communication and Technology

Paris, France
September 27-29, 1989

Probabilistic Multilevel Language Analysis for Speech Recognition

J. H. Wright (1), E. N. Wrigley (1), M. J. Carey (2)

(1) Department of Engineering Mathematics University of Bristol, UK (2) Ensigma Limited, Archway House, Welsh street, Chepstow, UK

Two adaptation methods for probabilistic context-free grammars are considered. In the first, an LR parser is enhanced with an error-recovery procedure which synthesises production rules in order to allow strings which are not in the language to be accepted. In the second, a hierarchy of probabilistic context-free grammars with progressively weaker structure has the same effect. Both methods are intended to promote flexibility and introduce a learning capability into linguistic modelling for speech recognition.

Full Paper

Bibliographic reference.  Wright, J. H. / Wrigley, E. N. / Carey, M. J. (1989): "Probabilistic multilevel language analysis for speech recognition", In EUROSPEECH-1989, 1486-1489.