MUSIC STRUCTURE DISCOVERY IN POPULAR MUSIC USING NON-NEGATIVE MATRIX FACTORIZATION
Florian Kaiser, Thomas Sikora · International Symposium/Conference on Music Information Retrieval · 2010
We introduce a method for the automatic extraction of musical structures in popular music. The proposed algorithm uses non-negative matrix factorization to segment regions of acoustically similar frames in a self-similarity matrix of the audio data. We show that over the dimensions of the NMF decomposition, structural parts can easily be modeled. Based on that observation, we introduce a clustering algorithm that can explain the structure of the whole music piece. The preliminary evaluation we report in the the paper shows very encouraging results.