摘要 |
PROBLEM TO BE SOLVED: To calculate a Cepstral Mean (CM) value for normalization, which cancels mismatch of feature quantity of input voice and a sound model, without producing response delay, irrelevant of voice recognition performance environment. SOLUTION: A background noise level measuring section 101 measures a background noise level L0. The background noise level L0 is accumulated in a storage section 102. A comparison section 103 compares a present background noise level L0, with a background noise level L1 measured at voice recognition start timing of last (previous) voice recognition, and accumulated in the storage section 102. A correction value calculation section 100 quantitatively grasps difference of performance environment of the present voice recognition from that of the last voice recognition, based on difference of background noise level. When change of the performance environment is a predetermined threshold or more, the CM value is calculated by a non-reflection mode in which the CM value calculated in the last voice recognition is not reflected to the present CM value. When the CM value is less than the predetermined threshold, the CM value is calculated by a reflection mode in which the CM value calculated by the last voice recognition is reflected to the present CM value. COPYRIGHT: (C)2011,JPO&INPIT
|