Comparison of conventional methods and deep belief networks for isolated word recognition

A comparative analysis of the use of conventional methods and Deep Belief Networks (DBN) for speaker independent Isolated Word Recognition on small vocabulary is discussed in this paper. The conventional methods of speech recognition include HMM/GMM framework and Multilayer Perceptrons (MLPs). Features from the speech frames are used to train MLPs using back-propagation. The features that are extracted are 12^th order LPCs and 39 dimensional MFCCs for each frame. The stacked Restricted Boltzmann Machines (RBM) constitute a Deep Belief Networks (DBNs). The DBN learning procedure undergoes a pre-training stage and a fine-tuning stage. DBNs gave a higher performance as compared with the conventional methods with an accuracy of approximately 93% for Isolated Word Recognition using MFCC features.