Uyghur Speech Recognition Based On Deep Neural Network

The general speech recognition are mainly used the combination of Hidden Markov Model and Gaussian Mixture Model to implement. But the hidden Markov model, after considering the triphone model,the scale of parameters increase fast, in the state of training data is limited, the model parameters are not well trained, thus affecting the speech recognition rate. In order to improve speech recognition rate, this paper proposes a method of speech recognition based on Uyghur deep neural network.In this paper, firstly, introduced the development situation at home and abroad of Speech recognition.introduced the the basic steps and principle of Speech recognition in detaily; secondly, describes the relevant principles and some shortcomings of hidden Markov model; then based on the the introduction Raises the deep neural network model focus of this article, Introduced the main principle and relevant algorithm BP algorithm and the RBM algorithm.finaly.according to the theory conducted experiments and Come to the conclusion.Deep neural network model presented in this paper is a high-capacity complex network model, it has many layers, each layer individually trained, the network model parameters are fully trained, especially graphing calculator(GPU) capabilities to effectively address the problem of large computation network. In this paper, kaldi as a test platform for training the uygher model based on the HMM-DNN, and using this model tests the Uyghur speech recognition. Experimental results show that the error in Uyghur speech recognition was dropped 2.72% after use the deep neural network model compared to the Subspace gaussian mixture model.It also reduced by 8.68% compared to HMM triphone. From the test results, DNN model can effectively reduce the Uyghur speech recognition error rate.the previous model optimization equally effective in uyghur deep neural network model.