Comparison of Partition-Based Audio Features for Music Classification

Inhalt:
Music classification is used to organize personalmusic collections or to provide recommendations for new songs and albums. This process is typically based on signal descriptors extracted from audio data which correspond to timbre, harmony, melody, rhythm and further structural characteristics. In this paper different audio feature sets ranging from a full set of low-level descriptors up to a more compact set based on cepstral properties are introduced. The classification task is to assign music tracks to 14 predefined music genre and style categories. A partition-based preprocessing is applied to all songs before they are categorized. For classification a small number of well-establishedmethods is used in comparison. The classification results show comparable accuracies for different feature sets in spite of significant differences in the number of features.