STATISTICAL SEQUENCE MODEL GENERATOR, STATISTICAL LANGUAGE MODEL GENERATOR, AND SPEECH RECOGNITION SYSTEM

摘要

PROBLEM TO BE SOLVED: To impart degree of freedom to an analyzed result, and to process a variable length of sequence as the same class, by re-estimating frequency probability of a bigram between sequences based on front likelihood, frequency probability, and rear likelihood. ;SOLUTION: Re-estimation is conducted in a statistical language model generating part 20 to provide the maximum likelihood estimate using EM algorithm based on frequency probability for character lines included in sorted classes and conditional character lines and frequency probability for a bigram between the classes. Frequency probability of a bigram between sequences is re-estimated using an expression for expressing the frequency probability of the bigram between the sequences based on front likelihood for the character lines of processing objects put in a front side of a time series, frequency probability of the character lines when the character line just before the character lines is conditioned, and rear likelihood corresponding to the character lines put in a rear side of the time series, relating to the respective character lines of the processing objects. Bi-multigram statistical sequence models are generated to be output.;COPYRIGHT: (C)1999,JPO