Phrase based feature extraction for musical information retrieval

This paper presents a feature extraction method for musical information retrieval. The method presented uses a feature vector consisting of frequencies of local patterns represented with n-grams, which enables efficient search from large databases. In order to improve the performance of music retrieval, we extract features at the phrase level instead of whole melody, using an automatic phrase extraction method. The presented method is applied to a small database of Japanese popular songs and its effectiveness is shown experimentally.