International Association of Engineers
Independent Component Analysis (ICA) is a powerful statistical method which can be used for many applications such as source separation, feature extraction or data representation. This paper proposes a new classification scheme spreading over two stages: density estimation using local ICA based on fuzzy clustering and correction of the estimation bias using SVM classification. The observed data are grouped into fuzzy clusters and linear ICA models are locally applied on each cluster. The classification experiments are carried out over multi-biometric feature vectors obtained from the fusion of lip movement and acoustic features.