Hybrid Model of Continuous Hidden Markov Model and Multi-Layer Perceptron in Speech Recognition

In order to overcome shortcomings of basic hidden Markov model (HMM), a hybrid model of multi-layer perceptron (MLP) and continuous hidden Markov model (CHMM) is presented which bases on basic HMM. In this hybrid mode, MLP calculates each state´s output probability instead of CHMM. The main purpose of this model is to improve the recognition ratio of CHMM by means of the strong of MLP´s nonlinear predictive capability. Speaker independent Mandarin digit speech recognition which based on the hybrid models is realized. Experimental results show that the hybrid model is efficiency and has higher recognition ratio than basic CHMM.