| 摘要 |
<P>PROBLEM TO BE SOLVED: To create a robust sound model from a small amount of training data, which is adapted for various regularization measures. Ž<P>SOLUTION: A device for sound model creation includes: a calculation section for statistic amount of correct answer; a prior distribution calculation section; a decoder; a calculation section for statistic amount of error; a calculation section for statistic amount of weighted difference; and a sound model update processing section. The calculation section for statistic amount of correct answer calculates a statistic amount of correct answer for each regularization measure. A prior distribution calculation section calculates and sets prior distribution according to the regularization measures. The decoder creates data for learning, which are easily mixed up, as error candidates. The calculation section for statistic amount of error calculates a statistic amount of error for each regularization measure. The calculation section for statistic amount of weighted difference calculates a statistic amount of the weighted difference of the statistic amount of correct answer and the statistic amount of error. Ž<P>COPYRIGHT: (C)2010,JPO&INPIT Ž
|
