Segmentation of Musical Signals Using Hidden Markov Models

In this paper, we present a segmentation algorithm for acoustic musical signals, using a hidden Markov model. Through unsupervised learning, we discover regions in the music that present steady statistical properties: textures. We investigate different front-ends for the system, and compare their performances. We then show that the obtained segmentation often translates a structure explained by musicology: chorus and verse, different instrumental sections, etc. Finally, we discuss the necessity of the HMM and conclude that an efficient segmentation of music is more than a static clustering and should make use of the dynamics of the data.