UCLA

Content-based music analysis has attracted considerable attention due to the rapidly growing digital music market. A number of specific functionalities, such as the exact look-up of melodies from an existing database or classification of music into well-known genres, can now be executed on a large scale, and are even available as consumer services from several well-known social media and mobile phone companies. In spite of these advances, robust representations of music that allow efficient execution of tasks, seemingly simple to many humans, such as identifying a cover song, that is, a new recording of an old song, or breaking up a song into its constituent structural parts, are yet to be invented. Motivated by this challenge, we introduce a method for determining approximate structural sequence representations purely from the chromagram of songs without adopting any prior knowledge from musicology. Each song is represented by a sequence of states of an underlying Hidden Markov Model, where each state may represent a property of a song, such as the harmony, chord, or melody. Then, by adapting different versions of the sequence alignment algorithms, the method is applied to the problems of: (i) Exploring and identifying repeating parts in a song; (ii) identifying cover songs; and (iii) extracting similar sections from two different songs. The proposed method has a number of advantages, including elimination of the unreliable beat estimation step and the capability to match parts of songs. The invariance of key transpositions among cover songs is achieved by cyclically rotating the chromatic domain of a chromagram. Our data-driven method is shown to be robust against the reordering, insertion, and deletion of sections of songs, and its performance is superior to that of other known methods for the cover song identification task.