Talker-Recognition Procedure Based on Analysis of Variance

Abstract

This paper describes a talker-recognition procedure that selects a small subset of features from a much larger body of data and bases recognition on this subset. The subset is chosen as those features having small variation between utterances of a given talker as compared to the variation among utterances of different talkers. This procedure is applied to materials consisting of quantized spectrographic information from a group of ten talkers uttering 10 different words seven times each; modified materials averaging the spectrographic information are also used for recognition. Results are computed as a function of the number of features used. These results are compared with those obtained by simple pattern matching and by a procedure that selects features by means of multidimensional analysis of variance. Although we have treated talker recognition only, this procedure is general and can be applied to other recognition problems where it is desirable to use a simple, cheap, and fast method to select a few important features from a much larger set that can be in the thousands.