Analytical and iterative approaches to the equalisation of sub-band errors in speech and speaker recognition

Recent work in both speech and speaker recognition has shown some interesting apparent benefits of sub-band processing: dividing the acoustic band into sub-units to give multiple stream (sub)-classifiers. In this paper we extend our recent work on sub-band error equalisation by considering 4 separate cases from combinations of male and female speaker sets in the context of speech and speaker recognition. We show that sub-band error equalisation can be achieved by changing the conventional mel frequency-warping function. This can also reduce the overall error rates significantly: in the case of the female speaker set the rate is reduced by over 50%.