Deep belief network optimization in speech recognition

The paradigm theory of neural network (NN) is expressing from biological of human brain function system. In this study, we deal with the deep belief network (DBN) model using the speech as natural language processing for training the system. The characteristic of a deep belief network is similar to a neural network. Whereas the performance of a neural network depends on the structure itself and it is suitable to select the model and size of the network for the data to handle. In the occasion, deep belief network has the advantage in speech recognition it generates the feature learning with a subsequent stage of supervised learning, once the network initialized, the first way using unsupervised and then fine-tuned with the labeled data to train some neuron in initial weight vectors. Deep belief network (DBN) have many nonlinear hidden layers to produce posterior of probabilities that take several frames of coefficient as input. This paper shows the good performance improvement of modularity in DBN as combination technique of computation.