To satisfy an error rate of a voice recognition device and to automatically construct a regression class tree structure by combining regression classes each other and forming regression classes more closely located to the root of the tree based on correlation parameters.
In the regression tree structure, a word subcluster indicating a square is used as a fundamental class. The word subcluster expresses a fundamental regression class and also a leaf of the tree structure. A joint of the tree is indicated by an ellipse and expresses a regression class combining two fundamental regression classes. Construction of a tree structure is started with these fundamental regression classes as starting points. In each step of the construction of the tree structure, two regression classes having a maximum correlation parameter are combined by a new tree joint to form a new regression class. The tree joint is closely located at the root of the tree than the two combined regression classes. By combining new regression classes, combining of two regression classes is continued until a regression class representing a tree root is formed.