Application Of Improved Hybrid Model Based On HMM And BP Neural Nework In Speech Recognition

Speech recognition is a technology which has rich content and has been widely used. This thesis focuses on the main issues of Chinese speech recognition. In order to improve the recognition ratio and speed up convergence, the key technologies of the Chinese speech recognition has been researched.This thesis analyzes the speech signal and describes the principle of speech recognition from the time domain, frequency domain and cepstrum domain. The filter banks analysis method and linear predictive coding technology are introduced, and the LPCC (Linear Predictive Coding Cepstrum) parameters and the MFCC (Mel-Frequency Cepstrum Coefficients) are given. In feature extraction, the MFCC parameters based on the human auditory model are chosen, and compared with the LPCC parameters based on the human phonation model.Hidden Markov Model (HMM) and Artificial Neural Network (ANN) have been widely used in speech recognition. Their respective advantages and disadvantages are analyzed and the combination methods of Hidden Markov Model and Artificial Neural Network are summarized. In view of the advantages and disadvantages of the hybrid models, a new hybrid approach is put forward. In the new method, the states output probabilities of Continuous Density Hidden Markov Model (CDHMM) are used as the input of the Back Propagation (BP) neural network. On the one hand, the BP neural network does not need a clear formula, because it can find out the intrinsic relationship between the input and the output by training and learning, according to the provided data. On the other hand, Discrete Hidden Markov Model (DHMM) will produce the quantization error, so the CDHMM/BP neural network hybrid model is adopted, for it takes advantage of the time domain modeling ability of the CDHMM and the strong classification ability of the BP neural network, and considers the characters of different classification sufficiently. Experimental results show that the hybrid method can improve the recognition ratio to a certain extent.This thesis summarizes the limitations of traditional BP training algorithm. Based on the former research, the commoner tan-sigmoid transform function is adopted. During the training, the zoom coefficients and the displacement parameters are adjusted with the weight matrix. The information is stored in weight...