Unsupervised Analysis and Generation of Audio Percussion Sequences

Abstract

A system is presented that learns the structure of an audio recording of a rhythmical percussion fragment in an unsupervised manner and that synthesizes musical variations from it. The procedure consists of 1) segmentation, 2) symbolization (feature extraction, clustering, sequence structure analysis, temporal alignment), and 3) synthesis. The symbolization step yields a sequence of event classes. Simultaneously, representations are maintained that cluster the events into few or many classes. Based on the most regular clustering level, a tempo estimation procedure is used to preserve the metrical structure in the generated sequence. Employing variable length Markov chains, the final synthesis is performed, recombining the audio material derived from the sample itself. Representations with different numbers of classes are used to trade off statistical significance (short context sequence, low clustering refinement) versus specificity (long context, high clustering refinement) of the generated sequence. For a broad variety of musical styles the musical characteristics of the original are preserved. At the same time, considerable variability is introduced in the generated sequence.