摘要 |
PROBLEM TO BE SOLVED: To conduct highly precise speech recognition while suppressing the increase in size of the resources that are required for processing. SOLUTION: A word preliminary selection section 13 selects one or more number of words following the words already obtained of the word string that becomes a candidate for speech recognition result. In a matching section 14, acoustic or linguistic score is computed for the selected word and a word string which becomes speech recognition result is constituted, based on the score. Then, in a control section 11, are word connecting relationship is generated among the words of the word string that becomes the candidate of speech recognition result, supplied to a word connecting information storage section 16 and stored. In a re-evaluation section 15, the word connecting relationship stored in the section 16 is successively corrected and in the section 11, a word string that becomes a speech recognition result is decided, based on the corrected word connecting relationship. |