Music thumbnailing incorporating harmony- and rhythm structure

A variety of approaches exist to the automatic retrieval of the key part within a musical piece - its thumbnail. Most of these however do not use adequate modeling with respect to either harmony or rhythm. In this work we therefore introduce thumbnailing that aims at adequate musical feature modeling. The rhythmic structure is extracted to obtain a segmentation based on beats and bars by an IIR comb-filter bank. Further, we extract chroma energy distribution normalized statistics features of the segmented song improving performance with dB(A) and pitch correction. Harmonic similarities are determined by construction and analysis of a similarity matrix based on the normalized scalar product of the feature vectors. Last, thumbnails are found lending techniques from image processing. Extensive test runs on roughly 24 h of music reveal the high effectiveness of our approach.