Reducing Dimensionality of Spectro-Temporal Data by Independent Component Analysis

Abstract

This paper studies the use of independent component analysis (ICA) for reducing the dimensionality of one type of spectro-temporal features, known as the MPEG-7 audio signature descriptors. The dimension-reduced features are used to identify distorted audio items in the experiments. The proposed ICA-based reduction approach is compared with the block average method and the principal component analysis (PCA) method. The experimental results show that features obtained by the ICA approach have higher identification accuracy than comparison counterparts for moderate to highly distorted soundtracks. In this regard, the proposed approach is a better alternative for dimensionality reduction for spectro-temporal features with distortion.