Classifier ensemble selection for language verification system

Spoken-language verification system uses classifier combination method to improve its performance. The number of classifiers combined determines the system's costs in time and calculation. Hence, we aim to get the optimal classifier ensemble with less cost and good performance. We hope to find some characteristics of classifier ensemble closely linked to its equal error rate (EER) and then choose the optimal classifier ensemble based on them. Two new diversity measures were proposed. Through rank correlation coefficients between them and EER, we found new diversity measures had closer correlation with the performance of system. Final results showed combining two new measures is the most effective to choose the optimal classifier ensemble, which makes system 14.71% best relative decrease in EER and about 60% best relative decrease in costs. We also explored preliminarily the robustness of this method over open-set corpus.