Automatic speaker identification for a large population

Design of speaker identification schemes for a small number of speakers (around 10) with a high degree of accuracy in a controlled environment is a practical proposition today. When the number of speakers is large (say, above 20 or 30), many of these schemes cannot be directly utilized as both recognition error and computation time increase monotonically with population size. A multistage classification technique gives better results when the number of speakers is large. Such a scheme may be implemented as a decision tree classifier in which the final decision is made only after a predetermined number of stages. In the present paper, analysis and design of a two-stage pattern classifier is considered. At the first stage a large number of classes, to which the given pattern cannot belong, is rejected. This is to be done using a subset of the total feature set. Also, the accuracy of such a rejection process must be very high, consistent with the overall accuracy desired. This initial classification gives a subset of the total classes, which has to be carefully considered at the next stage utilizing the remaining features for an absolute identification of the class label (the speaker's identity). The procedure is illustrated by designing and testing a two-stage classifier for speaker identification in a population of 30.