Three music structure estimation systems
The three systems produce an estimation of the semiotic structure of the music piece considered, i.e. a description of its macroscopic organization through a set of structural segments labeled according to the similarity of their musical content.
They consist in three steps: a feature extraction step, a segmentation step based on feature analysis under a regularity constraint, and a labeling step based on hierarchical clustering.
System 1 (2010) uses timbre homogeneity, tonal content repetitions and short sound events for segmentation. Resulting segments are clustered according to their timbre.
System 2 (2011) performs a segmentation through chord repetitions. Resulting segments are clustered according to the similarity of their chord sequence.
System 3 (2012) considers an internal model of the structural segments for segmentation. Resulting segments are clustered according to the similarity of their tonal content.
Authors: Gabriel Sargent, Frédéric Bimbot, Emmanuel Vincent
All: PC or Mac with Matlab (signal processing and statistics toolboxes)
System 1 (2010) requires the mfcc extractor from the MA Toolbox by Slaney and Logan, and chroma and beat extractors developed by Ellis (Coversongs project, LabRosa)
System 2 (2011) requires the chord estimation by Ueda (University of Tokyo), the beat and downbeat trackers by Davies (INESC Porto), and Matlab edit distance script by Miguel Castro (Matlab Central)
System 3 (2012) requires the Chroma Toolbox by Muller and Ewert (Max-Planck-Institut für Informatik) and the beat and downbeat trackers by Davies (INESC Porto)
The three systems have been developed at Irisa in Rennes and is the property of Université de Rennes 1, CNRS and Inria. They are currently prototypes provided by IRISA/PANAMA under the « Creative Commons Attribution-NonCommercial-ShareAlike 3.0″ license (http://creativecommons.org/licenses/by-nc-sa/3.0/legalcode)