| 摘要 |
PROBLEM TO BE SOLVED: To provide a speech recognition method where degradation in recognition accuracy due to estimation errors can be avoided, and also the composition of a clean voice model and a noise model being heretofore major cause for a delay can be performed before the input of a voice signal, and the considerable shortening of the processing delay is made possible. SOLUTION: Before the voice signal is inputted, the noise model is learned from at least the observed noise, and the composition of the clean speech model and the noise model and the calculation of the long-time average of the characteristic parameter of the reference signal superimposed with the observed noise are performed beforehand. When the voice signal is inputted, the characteristic parameter of the voice signal superimposed with the noise is extracted, and model collation likelihood calculation is performed by calculating the long-time average of the characteristic parameter. COPYRIGHT: (C)2006,JPO&NCIPI
|
