Independent component analysis for audio classification

In this paper, we explore the performance gains achieved by performing independent component analysis (ICA) decomposition on speech features obtained from a model of the early auditory system. ICA projection achieves dimensionality reduction by reducing the redundancy in the feature set (the transformed features are statistically independent). Performance is evaluated for an audio classification environment using a Gaussian mixture model (GMM) classifier and compared against the classification performance of AdaBoost, a wide-margin boosting algorithm. The new features are compared with mel-frequency cepstral coefficients (MFCC) and perceptual linear prediction (PLP) features. We also show that the ICA transformation is well suited for dimensionality reduction of auditory system-inspired features and it significantly improves the classification accuracy.