Visually-based audio texture segmentation for audio scene analysis

In an analogy with image texture segmentation in visual scene analysis, this paper describes a method for segmenting sound textures in a stream of audio. In particular, we propose a visual scheme to partition an audio stream signal into pieces of audio textures. This visual representation is based on the inter-similarity matrix of the MFCC feature in the signal frames. Classical image enhancement such as binarization and median filtering are applied to the inter-similarity matrix in-order to partition the matrix into homogenous regions. A novelty test operator is then used to localize the boundaries of the image regions, which correspond to audio textures boundaries, signalling thereby a change of audio scene. The perceptual and computational advantages of this visually-based audio texture segmentation are illustrated using a wide range of sound textures of varying degree of complexity.