STATISTICAL LANGUAGE MODEL GENERATING DEVICE AND VOICE RECOGNITION DEVICE

摘要

PROBLEM TO BE SOLVED: To minimize the number of parameters which are to be stored even when the language model is of trigrams or more and to obtain a high recognition rate. SOLUTION: A language model generating part 20 generates a statistical language model by smoothing and learning a statistical language model having the unigram of the word being a processing object, the bigrams between the word being the processing object and a word existing at a distance 1 immadiatey before the word, the bigrams between the word being the processing object and a word existing at a distance 2 immadiately before the word and model parameters including a normalization parameter normalizing the unigram and the two bigrams so that errors among the trigrams of the word being the processing object and estimation values of the two bigrams become minimum based on the text data for learning which are stored in a text data memory for leaning 21.