Audio Segmentation by Singular Value Clustering

Many audio texture and even musical pieces can be considered as an alternating juxtaposition of different sound types or objects. Determining segments in an audio signal that correspond to a coherent object (especially in unsupervised manner) is important for visualization purposes, audio texture synthesis and creative audio manipulations. This paper presents a statistical approach to sound texture modeling based on a singular value analysis of spectral features or an eigenvector analysis of their similarity matrix. We present a principled approach that brings methods such as audio similarity analysis and spectral audio basis representations into one framework.