Music Structure Discovery: Measuring the ”state-ness” of Times

Music Structure Discovery (MSD) aims at estimating the underlying structure of a music track using observations of the audio signal. For this, a given time ti of a music track is supposed to belong to one of the following categories: • if ti contains information similar to its adjacent times, ti is said homogeneous [7] and belongs to a ”state” [8], • if ti is similar to a foreign time tj , and if the same is true for ti+l l ∈ [1, λ] (similar to tj+l), we say that the corresponding segments are repetitions [7]. If the corresponding times ti+l are similar to their adjacent times than we have a ”state repetition”. If this is not the case, we say that the times ti+l and tj+l belong to a ”sequence” [8] of length λ which is instantiated at time ti and at time tj . • if ti is not similar to any other times, ti is a null-time. This subdivision has lead to two types of approaches to estimate the music structure: • the state approach, which is used to detect states (being repeated or not) and • the sequence approach, which is used to detect sequences (i.e. repetitions which are not states). See [8] for more details. This subdivision if summarized in the table below.