Biometric Authentication: A Machine Learning Approach

9.3. Kernel-Based Probabilistic Speaker Models

Because the amount of speaker-dependent data in a speaker recognition task is typically very large, it is almost impossible to store all data in the form of templates for recognition. Early techniques, such as vector quantization [342], attempt to reduce the amount of data by replacing similar data with their corresponding centroids. This is equivalent to partitioning the feature space into a number of clusters. This technique, however, assumes that data falling on one cluster do not influence the other clusters. In recent years, a number of kernel-based probabilistic neural networks have been proposed to address the deficiency of VQ; they include Gaussian mixture models (GMMs) [307], elliptical ...