To create an acoustic model having a high recognition rate irrespective of a quantity of learning data.
The method for creating an acoustic model for speech recognition is established by providing: the acoustic model with a primary acoustic model storage part in which a primary acoustic model is stored; and a learning data storage part in which learning data are stored. The method obtains a movement vector of a Gauss distribution average vector of the acoustic model from the learning data acquired from the primary acoustic model and the learning data storage part, and transforms the primary acoustic model into an adaptive acoustic model adaptive to the above learning data by using this movement vector. The movement vector is decomposed into a direction vector and a scaling factor, the direction vector is obtained by learning in parameter estimation of a coarse class, and the above scaling factor is obtained by learning in parameter estimation of a fine class.
COPYRIGHT: (C)2006,JPO&NCIPI
渡部 晋治
中村 幸雄
稲垣 稔
草野 卓