Automatic Prediction of Hit Songs

We explore the automatic analysis of music to identify likely hit songs. We extract both acoustic and lyric in- formation from each song and separate hits from non-hits using standard classiers, specically Support Vector Ma- chines and boosting classiers. 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 per- formance that is much better than random. The lyric- based features are slightly more useful than the acoustic features in correctly identifying hit songs. Concatenat- ing the two features does not produce signicant improve- ments. 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.