The Appropriate Hidden Layers of Deep Belief Networks for Speech Recognition

Abstract

Recently, Deep Belief Networks (DBNs) have received much attention in speech recognition communities. However, there are rare methods to set the appropriate hidden layers of DBNs. In this paper, we study the relationship between the number of hidden layers and the invariant features of speech signals, and the time cost of the accuracy of speech recognition. Also, we study the approximations in Contrastive Divergence algorithm which is used to train the Restricted Boltzmann Machine. We conclude that it exists an appropriate number of hidden layers of DBNs which can balance the accuracy of speech recognition and the training time. It has appropriate number of hidden layers of DBNs for the experiments of speech recognition on TIMIT corpus. When the number of hidden layers greater than the appropriate number the accuracy of speech recognition are almost the same, and the time cost increase largely.