Pair-Wise Discriminant Functions

T.Kawahara et al (1993) presented the Pair-Wise discrimination method that uses Bayes classifiers based on discriminant functions. First, a set of pairs of categories is constructed. The number of the pairs is possibly , where n is the number of categories. However, as the number of categories increases, all the possible combinations drastically increases. Therefore, the effective pairs for overall discrimination are selected, along with the measurement of discrimination ability. Next, for each pair of the categories, the discrimination probability is calculated with its corresponding discriminant function, and finally the classified result is obtained by the operation such as minimax, average, and majority. A detailed illustration is shown in Figure 2.10. Pair-Wise classifier is very effective to discriminate local patterns statistically between similar pairs of categories, e.g., fricatives and plosives, fricatives and affricates, etc.

  
図 2.10: Pair-Wise discrimination method