By Jean-Julien Aucouturier, François Pachet, Pierre Roy, and Anthony Beurivé.
In the Proceedings of the 8th International Conference on Music Information Retrieval (ISMIR 2007), pp. 425–430, Vienna, Austria, 2007.
Read the paper (PDF)
ISMIR proceedings PDF
About the paper
A track’s genre, instruments, and mood are often related. This paper explores how those relationships can help an automatic music tagger make better predictions from audio.
Abstract summary
The system starts with acoustic classification and uses its predictions as inputs to decision trees. The trees learn relationships among metadata labels, and the process repeats to refine the predictions. The experiments examine the gains from this combination on musical descriptions, including cultural and subjective categories.
Related work
Improving Multilabel Analysis of Music Titles: A Large-Scale Validation of the Correction Approach · PDF (2009), by François Pachet and Pierre Roy, evaluates the correction approach on a larger catalog.
See Timbre Similarity and Hubness for the audio models, Feature Generation for work on task-specific descriptors, and Active Listening for applications to music browsing.