Isolated Word Speech Recognition System Using Deep Neural Networks

Speech recognition is the process of converting speech signals into words. For acoustic modeling HMM-GMM is used for many years. For GMM, it requires assumptions near the data distribution for calculating probabilities. For removing this limitation, GMM is replaced by DNN in acoustic model. Deep neural networks are the feed forward neural networks having more than one or multiple layers of hidden units. In this work, we have presented the isolated word speech recognition system using acoustic model of HMM and DNN. We are using Deep Belief Network pre-training algorithm for initializing deep neural networks. DBN is a multilayer generative probabilistic model with large number of stochastic binary units. The features used are the mel-frequency cepstrum coefficients (MFCC). Experimental results are calculated on TI digits database. Proposed system has achieved 86.06 % accuracy on TI digits database. System accuracy can be further increased by increasing the number of hidden units.