Genre Classification for Musical Documents Based on Extracted Melodic Patterns and Clustering

Genre classification for musical documents is conventionally based on keywords, statistical features or low-level acoustic features. Such features are either lack of in-depth information of music content or incomprehensible for music professionals. This paper proposed a classification scheme based on the correlation analysis of the melodic patterns extracted from music documents. The extracted patterns can be further clustered, and smoothing techniques for the statistics of the patterns can be utilized to improve the performance effectively. The accuracy of 70.67% for classifying five types of genre, including jazz, lyric, rock, classical and others, can be achieved, which outperforms an ANN-based classifier using statistical features significantly. The patterns can be converted into symbolic forms such that the classification results are meaningful and comprehensible for most music workers.