Unsupervised Generation of Percussion Sequences from a Sound Example

In this masters project, a system is developed for the analysis of the structure and the style of a percussive audio sequence with the aim of generating an arbitrarily long musically meaningful and interesting sound sequence with the same stylistic characteristics as the original. The framework is developed around a Variable Length Markov Chain (VLMC) that is used to predict and re-shuffle the basic bricks of a musical phrase, the musical events. Using an onset detector, at first, the sound is split into segments. Applying various clustering thresholds simultaneously, the clustering of the segments yields a multi-level discrete representation of the sound sequence. Then a regularity estimation of the levels is performed to determine the levels of refinement where regular time patterns appear. Those periodic events are then used as the core structure on top of which a regular temporal grid is built. The temporal grid defines the events, the tempo, and the meter. Finally different resolution levels are taken into account for the prediction using a generation strategy enhancing statistical significance while maximizing cluster resolution in order to avoid musical discontinuities. On a data base of percussion and beat boxing examples, the system was evaluated objectively as well as subjectively by two professional percussionists, showing that the automatically generated sequences maintain the style and beat of the original and that they are musically interesting.