Text Independent Speaker Recognition Using A Fuzzy Hypercube Classifier.

The recognition of speakers in an open set, text-independent environment is described. The recognition occurs without any prior training, and occurred in both noisy and clear backgrounds in as little as 1.6 seconds. Investigations and testing were done in the areas of feature characterization of speakers, prefiltering of classifier input, and structure of classifiers for recognition. A prefiltering structure for speech input segments using an expert system implementing hypothesize and test for relevance was investigated. This attempts to maximize classification performance by preselection of most likely voiced speech segments prior to classification. The classifier used was based on Adaptive Resonant Theory and fuzzy Min-Max. It is a neural network with output categories represented by a fuzzy hypercube. The network is described in a hybrid neuronal-functional method. A speaker recognition system was tested using the Switch-board and Greenflag databases. Utterances averaging 0.5 to 7.0 seconds in length were tested, with over 5 hours of conversation for 8, 12 and 16 speaker groups.