To provide a technology for reducing the size of an acoustic model without additional learning.
An acoustic model generation device includes an acoustic model storage section 13 for storing the acoustic model composed of a plurality of mixed normal distributions, which are composed by weighted addition of the plurality of base normal distributions respectively. A base normal distribution deleting section 162 deletes a base normal distribution to which a weight smaller than a predetermined deletion weight threshold is added, from the plurality of base normal distributions for composing the mixed normal distribution, for each of the whole or a part of mixed normal distribution for composing the acoustic model. A normalization section 17 normalizes each of the mixed normal distributions from which the base normal distribution is deleted in the mixed normal distributions for composing the acoustic model so that the sum of the weight added to each of the plurality of base normal distributions for composing the mixed normal distribution, which remains without being deleted, can become one.
COPYRIGHT: (C)2009,JPO&INPIT
Atsutoku Ogawa
| JP6259089A | ||||
| JP2005321660A | ||||
| JP2008233759A | ||||
| JP2004272052A | ||||
| JP2001022368A |
Yukio Nakamura
Yoshimura Munehiro
Taku Kusano