A Novel Framework Based on Trace Norm Minimization for Audio Event Detection

Abstract

In this paper, a novel framework based on trace norm minimization for audio event detection is proposed. In the framework, both the feature extraction and pattern classifier are made by solving corresponding convex optimization problem with trace norm regularization or under trace norm constraint. For feature extraction, robust principle component analysis (robust PCA) via minimizing a combination of the nuclear norm and the l1-norm is used to extract matrix representation features which is robust to outliers and gross corruption for audio segments. These matrix representation features are fed to a linear classifier where the weight matrix and bias are learned by solving similar trace norm regularized problems. Experiments on real data sets indicate that this novel framework is effective and noise robust.