Music Structure Analysis


The task of music structure analysis (also known as music segmentation) is a relevant one in the field of Music Informatics. The main goal is to automatically segment a specific piece of music into its various sections, and group them based on their acoustic similarity (so that we have section A, B, C, or verse, bridge, chorus, etc).


In the lab, we have proposed various methods of music structure analysis:

Shift-Invariant Probabilistic Latent Component Analysis

Music structure segmentation based on the probabilistic approach of the matrix factorization machine learning technique called SI-PLCA (Shift Invariant Probabilistic Latent Component Analysis), using beat-synchronous chroma features.

Please, download the source code from github.

Convex Non-negative Matrix Factorization

Difference between C-NMF and traditional NMF for the song

Exploring various matrix factorization techniques for the automatic music structure analysis of pop songs. In this case, we made a comparison between Non-negative Matrix Factorization (NMF) and Convex-NMF. C-NMF yields better results when clustering the segments on the Beatles dataset and on the SALAMI dataset.


Music Summaries

We proposed a criterion to find the segments that best represent a musical piece, and concatenate them in order to create an audio summary of the track. This criterion balances two main important measures: how well the segments can compress the entire track (repetitiveness), and he amount of overlap between segments (mutual information).

Various experiments were performed on the Chopin’s Mazurkas dataset with encouraging results.


This material is based upon work supported by the NSF (grant IIS-0844654), by the IMLS (grant LG-06-08-0073-08), and by Fundación Caja Madrid.