LANGUAGE MODEL CREATION DEVICE, METHOD AND PROGRAM THEREOF

摘要

PROBLEM TO BE SOLVED: To provide a language model creation device capable of using a latent language model having an excellent language prediction performance for real time voice recognition processing.SOLUTION: An LWLM learning section inputs a learning text to generate plural latent strings which are latent word strings corresponding to the respective word strings of the learning texts to learn a latent language model including two probability distributions of a latent word-latent word probability and a latent word-observed word probability from the latent strings. A pseudo learning text creation section inputs the latent language model generated by the LWLM learning section and generates a latent word string from the latent word-latent word probability to create a pseudo text from the latent word string and the latent word-observed word probability. An N-gram language model creation section inputs the pseudo text created by the pseudo learning text creation section and counts the frequency of the words all of N-sets in the pseudo text to create an N-gram language model.