Audio Segmentation by Feature-space Clustering Using Linear Discriminant Analysis and Dynamic Programming

We consider the problem of segmenting an audio signal into characteristic regions based on feature-set similarities. In the proposed method, a feature-space representation of the signal is generated; then, sequences of feature-space samples are aggregated into clusters corresponding to distinct signal regions. The clustering of feature sets is improved via linear discriminant analysis (LDA); dynamic programming (DP) is used to derive optimal cluster boundaries. The method avoids the heuristics employed in various feature-space segmentation schemes and is able to derive an optimal segmentation once the LDA and DP cost metrics have been chosen. We demonstrate that the method outperforms typical feature-space approaches described in the literature. We focus on an illustrative example of the basic segmentation task; however, by judicious design of the feature set, the training set, and the dynamic program, the method can be tailored for various applications such as speech / music discrimination, segmentation of audio streams for smart transport, or song structure analysis for thumbnailing.