| 摘要 |
PROBLEM TO BE SOLVED: To improve a sentence recognition rate of speech recognition by using a large-scale corpus and a less-restricted model like an N-gram model and to deter a model size and a search space from becoming enormous. SOLUTION: A speech recognition device includes a speech recognition part 44 which performs speech recognition based upon an N-gram language model 41, a syllable number calculation part 46 which estimates the number of syllables in an input speech from the recognition result, an FSA part 48 which stores FSAs 60 generated from subsets obtained by classifying sentences in a corpus by the numbers of syllables, a selection part 50 which selects 11 FSAs 52, i.e. the number of syllables that the syllable number calculation part 48 estimates and five corresponding FSAs 52 each before and after it among the FSAs 60, a speech recognition part 54 which recognizes the input speech according to the selected FSAs 52, and a selection part 56 which selects one of them according to acoustic scores of speech recognition results of the speech recognition parts 44 and 54. COPYRIGHT: (C)2004,JPO&NCIPI |
