In this thesis, we proposed 2nd Markov speech recognition structure which combines HMM with Neural Networks. We applied this structure to Korean digit recognition problem and compared the result with that of existing HMM structure. We compensated HMM structure for 2nd Order Markov property using Neural Network. By using the proposed structure, we can use previous observations as inputs of Neural Network and also, we obtained better recognition rates. For the proposed structure, we obain the recognition rates of 95.9% and it is better than that of existing HMM, 94.7%. The simulation results are represented.
