shrinking FSA for speech recognition

Hi,


      I'm working on language modelisation for speech recognition. We
want to compile some grammars similar to EBNF into FSA or word graphs
(in which labels appear in nodes instead of on edges).


Unfortunately, this compilation can generate very large FSA. Actually,
medium to large size grammars generate far too large FSA. So we would
like to transform those FSA into much smaller ones, approximating our
language model, without significantly reducing recognition accuracy.


Does someone know of any work that has been done in that direction ?


thanks in advance,


Francis Picard
M.Sc. student
DIRO, University of Montréal


picardfr@iro.umontreal.ca

Return to the comp.compilers page.
Search the comp.compilers archives again.