Abstract
This paper proposes a new calculation method of distance between music pieces based on personal preferences. In the proposed method, in order to utilize musical features for the distance calculation, feature vectors which represent melody, rhythm, and timbre, are firstly calculated from traning data sets. Then, by using the feature vectors, a metric having a general form of the Mahalanobis distance is defined and learned. Specifically, the metric is learned by minimizing a cost function designed so that the cost increases when the distance between dissimilar music pieces is shorter than the distance between similar music pieces. According to the distance based on the metric, the accurate retrieval of similar music can be expected.
