Modeling Music as a Dynamic Texture

摘要: We consider representing a short temporal fragment of musical audio as dynamic texture, model both the timbral and rhythmical qualities sound, two important aspects required for automatic music analysis. The texture treats sequence feature vectors sample from linear dynamical system. apply this new representation to task song segmentation. In particular, we cluster fragments, extracted song, samples mixture (DTM) model. show that DTM can accurately coherent segments in detect transition boundaries. Moreover, generative character proposed makes it amenable wide range applications besides As examples, use models songs suggest possible improvements other information retrieval such annotation similarity.