摘要 |
PROBLEM TO BE SOLVED: To provide a voice recognition system capable of acquiring an accurate recognition result candidate during utterance. SOLUTION: A voice detection part 2 starts cutting out a voice signal with a shift ofΔT in each frame length T. An acoustic analysis part 3 extracts an acoustic parameter representing the features of the voice signal from the cut-out voice signal. The 1st and 2nd collation parts 4, 7 collate the acoustic parameter from the beginning up to the middle of the utterance with a recognition vocabulary standard pattern 5 and a background pattern 8, respectively, and calculate likelihood of phonemic trains constituting the recognition vocabulary. A recognition result candidate determination means 10 consists of a normalized score 11 and a recognition result candidate determination part 12, and selects a phonemic train with the highest likelihood among the phonemic trains obtained from the 1st and 2nd collation parts 4, 7, and determines only a phonemic train as a recognition result candidate that is very likely to be the phonemic train of the input voice from the beginning up to the middle of the utterance based on the likelihood of each phonemic train. COPYRIGHT: (C)2006,JPO&NCIPI
|