Content-Based Audio Classification with Generalized Ellipsoid Distance

While the size of multimedia database increases, the demand for efficient search for multimedia data becomes more and more urgent. Most recent works on audio classification and retrieval adopt Euclidean distance as their distance measures. However, Euclidean distance is not a perceptual distance measure for some audio features. The purpose of this work is to derive two new distance measures for content-based audio classification, which are based on re-weighting and de-correlating each feature. Weighted Euclidean distance uses a diagonal matrix, which re-weighs the importance of each feature, and generalized ellipsoid distance takes further consideration on correlation between any two features. An audio database of 85 sound clips is used as our training set. The experimental results show that the generalized ellipsoid distance yields the best result and achieves an overall correction rate of classification.