Music auto-tagging with variable feature sets and probabilistic annotation

This paper proposes a music auto-tagging system based on probabilistic annotation of semantically meaningful tags with variable feature sets. The perception-related long-term features are extracted. The original features are selected by a combination algorithm of ReliefF and principle component analysis (PCA) to form a variable unique feature subset for each tag. The Gaussian mixture models (GMMs) are then trained for each tag. The test tracks are tagged by the output probability of GMMs. To evaluate the quality of the proposed music auto-tagging system, the per-tag precision and recall rates and F-score are measured. Experiment results indicate that the performance of the models trained with the original feature sets is comparable with those trained with MFCC. The reduced variable feature sets demonstrates 2% and 5% up than the original system in precision and recall rates.