An HMM/MFNN hybrid architecture based on stacked generalization for speaker identification

A hybrid architecture based upon hidden Markov models (HMMs) and multilayer feedforward neural network (MFNN) is presented for speaker identification. Unlike most of the previous combing methods, the proposed architecture uses HMMs to model individual speaker and uses MFNN to deal with the inter-speaker information for improving performance. Learning in the proposed architecture consists of two phases and, in particular, only a small amount of data is needed for the training. The HMM/MFNN architecture has been applied to text-independent speaker identification. Simulation has shown that the hybrid architecture yields better identifying rate than that of the conventional methods and other hybrid architectures