Automatic Prediction of Hit Songs

We explore the automatic analysis of music to identify likely hit songs. We extract both acoustic and lyric information from each song and separate hits from non-hits using standard classifiers, specifically Support Vector Machines and boosting classifiers. Our features are based on global sounds learnt in an unsupervised fashion from acoustic data or global topics learnt from a lyrics database. Experiments on a corpus of 1700 songs demonstrate performance that is much better than random. The lyricbased features are slightly more useful than the acoustic features in correctly identifying hit songs. Concatenating the two features does not produce significant improvements. Analysis of the lyric-based features shows that the absence of certain semantic information indicates that a song is more likely to be a hit.