To generate a weighted finite state transducer even though a large K is used.
Conversion information N of a weighted finite state transducer WFST which generates information required for voice recognition being inputted to a recognition section that is incorporated in a voice recognition device is realized by N=opt(optHC*opt(L*G)). While describing a state transition process from a state number group and when the state number group is the same, one state transition process is described and when an overlapped phoneme column exists in a portion of the category which is received in two state transition processes, the two state transition processes are connected and a process is provided to a phoneme group while phoneme environment is eliminated.
HORI TAKAAKI
Minoru Inagaki