Multiple hypotheses at multiple scales for audio novelty computation within music

摘要: Novelty-based segmentation of audio signals has proven good performances for the estimation boundaries structural sections within music pieces. However, are detected only if satisfy condition sufficient acoustic inner-homogeneity. While this constraint is very restrictive and not representative all musical contents, we propose in paper to extend detection novelty transitions between homogeneous non-homogeneous vice versa. Moreover, length considered boundary crucial, also introduce a multi-scale approach that allows capture different temporal scales same segmentation. Evaluation combination these two methods proves convincing results Embedding algorithm structure system, show can be consistently improved task.