REAL-TIME UNSUPERVISED MUSIC STRUCTURAL SEGMENTATION USING DYNAMIC DESCRIPTORS

Publication Type:

Conference Paper

Source:

SMC Conference 2011 (2011)

Abstract:

This paper presents three approaches for music structural segmentation, i.e. intertwined music segmentation and labelling, using real-time techniques based solely on dynamic sound descriptors, without any training data. The first method is based on tracking peaks of a sequence obtained from a weighted off-diagonal section of a dissimilarity matrix, and uses Gaussian models for labelling sections. The second approach is a multi-pass method using Hidden Markov Models (HMM) with Gaussian Mixture Models (GMM) in each state. The third is a novel approach based on an adaptiveHMMthat dynamically identifies and labels sections, and also sporadically reevaluates the segmentation and labelling, allowing redefinition of past sections based on recent and immediate past information. Finally, a method to evaluate results is presented, that allows penalization both of incorrect section boundaries and of incorrect number of detected segments, if so desired. Computational results are presented and analysed both from quantitative and qualitative points-of-view.

AttachmentSize
smc2011_submission_144.pdf1.29 MB
SMC paper: 
SMC BIBLIOGRAPHY