| 摘要 |
<p><P>PROBLEM TO BE SOLVED: To reduce recognition errors of words or word strings having low frequencies in appearance in sentence examples for study of a language model in voice recognition. <P>SOLUTION: A voice recognition device which converts an input voice 1001 uttered by a user into feature vectors by a voice analysis means 2001, converts them into a syllable string having a maximum likelihood by a syllable string recognition means 3001, and refers to statistics of word chains stored in a language model 4003 to convert the syllable string into a word string by a word string search means 4001 is provided with a short word language model 4004 where statistics of isolated words are stored and a utterance length estimating means 4005 which estimates the number of syllables included in the syllable string outputted from the syllable string recognition means 3001 and switches a language model which the word string search means 4001 should refers to, from the language model 4003 to the short word language model 4004 by the language model switching means 4006 in the case of a short utterance input, and thus recognition errors caused by short words or utterance re-input of word strings in correction work or the like are reduced. <P>COPYRIGHT: (C)2004,JPO</p> |
