Application Of Improved GRBM In Speech Recognition

Speech recognition is an important research direction of artificial intelligence and various speech recognition models have been developed, such as Hidden Markov Model(HMM), Dynamic Time Warping(DTW) etc.. Nowadays, speech recognition is widely applied in various fields of daily life, such as the mobile terminal, smart homes, security systems and many more.Deep learning neural network(DNN) is a hot topic of machine learning in recent years. The application of DNN and other improved algorithm in the speech recognition system is an important issue in the field. In the present paper, the Gaussian‐Bernoulli restricted Boltzmann machine(GRBM) is used to train and recognize the speech signal basing on a developed recognition method. The principle of training methods, theoretical models, and advantages of GRBM are reviewed. We proposed two schemes to improve the model of Gauss restricted Boltzmann machine. Firstly, by combining with the Parallel Tempering learning algorithm, an improved GRBM network based on Parallel Tempering(GRBM‐PT)is proposed to improve the sampling algorithm of GRBM. Secondly, the training algorithm of model parameters is improved. This is realized by using an improved traditional contrastive divergence(CD) algorithm based on Exponential Moving Average(EMA) to train the GRBM model. The parameters of GRBM were updated by considering all the model parameters got since the beginning of training. The improved algorithm was then realized in a developed experimental platform. The experimental platform of speech recognition system consists of two parts: the MATLAB platform and the embedded ARM9 platform. The MATLAB platform is used for simulation. The simulation is realized by combining the improved Gauss restricted Boltzmann machine with the support vector machine(SVM) classifier. The test result is found by the accuracy and convergence curve got in the simulation. The embedded ARM9 platform is used for the realization of the improved Gauss restricted Boltzmann machine. By combining with logic regression(LR) classifier, the proposed system was transplanted to the embedded ARM9 platform and the test was completed. Experimental results of digit speech recognition show that the improved model has a good performance in GRBM, especially the accuracy of speech recognition is greatly improved.