Extraction of characteristic music textures (eigen-textures) via graph spectra and eigenclusters

There exists a great diversity in the area of automatic audio segmentation since audio can be segmented based on various desirable aspects. However, the instances of texture change are not equally as important for all applications than the texture itself. Typically, audio can contain a variety of textures and some of them are often repeating. Thus, only the texture change instant is not sufficient for complete characterization of given audio since it lacks the ability to judge similar textures discontinuous in time. The accurate identification of characteristic textures is crucial for many applications like classification, indexing, browsing and summarization. In this paper, graph spectra and graph eigenclusters are proposed as a scalable technique for extracting predominant textures or eigentextures in a given musical audio and has yielded encouraging results. This approach not only makes segmentation more tractable and scalable but also helps in modeling given audio in terms of graphical structure, which is more perceptually revealing.