Musical Typicality: How Many Similar Songs Exist?

We propose a method for estimating the musical “typicality” of a song from an information theoretic perspective. While musical similarity compares just two songs, musical typicality quantifies how many of the songs in a set are similar. It can be used not only to express the uniqueness of a song but also to recommend one that is representative of a set. Building on the type theory in information theory (Cover & Thomas 2006), we use a Bayesian generative model of musical features and compute the typicality of a song as the sum of the probabilities of the songs that share the type of the given song. To evaluate estimated results, we focused on vocal timbre which can be evaluated quantitatively by using the singer’s gender. Estimated typicality is evaluated against the Pearson correlation coefficient between the computed typicality and the ratio of the number of male singers to female singers of a song-set. Our result shows that the proposed measure works more effectively to estimate musical typicality than the previous model based simply on generative probabilities.