Speaker identification based on the time-delay hierarchical mixture of experts

In this paper, we explore the hierarchical mixture of experts (HME) architecture for a substantial problem, that of text-dependent speaker identification. For a specific multiway classification, we propose a generalized Bernolli density instead of the multinomial logit density. Time-delay technique is also introduced to HME for spatio-temporal processing. Using the proposed density and the time-delay HME along with the EM algorithm, we show that the system has a satisfactory performance and yields significantly fast training