"Name That Song!" A Probabilistic Approach to Querying on Music and Text

We present a novel, flexible statistical approach for modelling music and text jointly. The approach is based on multi-modal mixture models and maximum a posteriori estimation using EM. The learned models can be used to browse databases with documents containing music and text, to search for music using queries consisting of music and text (lyrics and other contextual information), to annotate text documents with music, and to automatically recommend or identify similar songs.