| 摘要 |
<P>PROBLEM TO BE SOLVED: To provide an acoustic model which is not restricted by a condition that the SN ratio of an input voice is known. <P>SOLUTION: A Gaussian mixture model generation part 11 generates a multiple mixed Gaussian mixture model in one state on the basis of the waveform signal data of a plurality of kinds of environmental noise for learning which are stored in a database 21, so that an output likelihood can be maximum, and an HMM synthesizing part 13 generates a plurality of adapted HMMs which include mixture Gaussian distributions in respective states, which are represented by the sum of linear coupling of Gaussian distributions weighted by a prescribed weight coefficient in all combined states of respective states and correspond to a plurality of SNs between the noiseless voice HMM and the generated noise Gaussian mixture models, in accordance with a prescribed noiseless voice HMM and the generated noise Gaussian mixture models, and the plurality of generated adapted HMMs are juxtaposed to generate an acoustic model in a multipath form. A voice recognition part 4 uses the adapted acoustic model to perform voice recognition of an uttered voice signal on the basis of an extracted feature quantity. <P>COPYRIGHT: (C)2003,JPO |
