A pattern recognition technique is described in which a parametric representation of input signals or stimuli is employed. An input is considered as a vector, while the stimulus class is a multivariate process in the vector space. An adaptive sample set construction technique is described through which the conditional joint probability density of a class is approximated by the sum of Gaussian densities. The mean of each such density is an adaptively chosen "typical" sample of the class, and the set of samples so chosen are contained in the region of the space in which samples of the class are most populous. The decision process using the typical samples partitions the space into regions that envelop the chosen samples of a class. Arbitrary shaped and multiply connected regions can be constructed in this way, and multimodal probability densities can be approximated with a computationally simple procedure. Decision making on an incomplete set of parameters and on multiple observations of the input stimulus are discussed. This technique was successfully applied to the automatic recognition of speaker identity regardless of the spoken test. Experimental results are given.