Pattern recognition device, pattern recognition method and computer program product

摘要

A pattern recognition device includes a feature vector calculator; a model selecting unit; a correction vector calculator; a feature vector correcting unit; and a pattern recognition unit. The correction vector calculator calculates, for each of the selection models, a modified directional vector having N dimensional components (N≧1). A value of the n-th dimensional component (N≧n≧1) of the modified directional vector is obtained by subtracting a value of the n-th dimensional component of the variance vector multiplied by a predetermined coefficient from an absolute value of the n-th component of a different vector between the average vector and feature vectors to obtain a first value, and then multiplying the first value by a plus or minus sign identical to a sign of the n-th dimensional component of the difference vector, and further calculate a correction vector with respect to a vector obtained by superimposing the modified directional vectors.

主权项

1. A pattern recognition device comprising:
a feature vector calculator configured to calculate a feature vector from an input pattern including speech data, character data, or image data, the feature vector being used for speech recognition, character recognition, or image recognition; a model selecting unit configured to select, from models, one or more selection models whose distances from the feature vector are within a predetermined range, each of the models being represented by both an average vector of a pattern and a variance vector indicating variance of the pattern, and the average vector and the variance vector each having N dimensional components (N≧1); a correction vector calculator configured to calculate a correction vector based on a modified directional vector having the N dimensional components, a value of an n-th dimensional component (N≧n≧1) of the modified directional vector being an average value of one or more modified values, each of the modified values being obtained for one of the selection models and being obtained by multiplying a first value and a second value when the first value is greater than zero and being zero when the first value is less than or equal to zero, the first value being obtained by subtracting a value of the n-th dimensional component of the variance vector multiplied by a predetermined coefficient from a difference, and the second value being obtained by dividing a third value by the difference, wherein the third value is obtained by subtracting a value of the n-th dimensional component of the average vector from a value of the n-th dimensional component of the feature vector, and wherein the difference is a difference between the value of the n-th dimensional component of the average vector and the value of the n-th dimensional component of the feature vector; a feature vector correcting unit configured to generate a corrected feature vector based on the correction vector and the feature vector; and a pattern recognition unit configured to perform the speech recognition, the character recognition or the image recognition by using the corrected feature vector.