Speech Recognition Based On HMM And Improved Neural Network

In the recent years,artificial intelligence and human-computer interaction are playing diversified role in many applications.There is necessity and demand to make the machines to recognize and understand human languages,so that the users around the world feel connected to the human voice in different languages.In the recent years,with the rapid development of modern voice recognition technologies,human voice has become one the key ways for human-computer interaction.This led the speech recognition technology as a widespread research topic in multiple disciplines.Considering these developments,in this thesis,Hidden Markov(HMM)and improved neural network speech recognition studied extensively.The main work accomplished can be summarizes as follows.Firstly,the steps involved in preprocessing of speech is analyzed and studied.Initially,the recorded analog voice signal is correlated and digitized to obtain its corresponding digital signal.Then the signal is weighted to facilitate the external influence of noise generation,and the high-frequency components contained in the speech are strengthened,which lays a firm foundation for the study of related parameters of the subsequent speech signals.Afterwards,a window function is attached to the speech signal,and then each fixed speech signal is divided into a plurality of small-segment speech forms in a frame as a basic unit,so as to facilitate subsequent operation processing.Finally,endpoint detection is performed on the speech signal to determine the start and end points of the speech signal.In the second step,the feature extraction algorithm of speech signal is analyzed and studied.The feature parameter sequences capable of characterizing speech features are separated and extracted,which helps to improve the evaluation and training of HMM for speech signals,which is beneficial to the classification and training of neural networks.In this thesis,the algorithm that can extract the characteristic parameter values related to speech signals is deeply analyzed,and the algorithm of using Mel-Frequency Cepstral Coefficients to obtain feature parameters is studied,and the specific process of designing and implementing the algorithm is explored.The overall architecture design flow of speech recognition and design of recognition algorithms are introduced as a third step,and the recognition algorithms commonly used so far are described.In view of the analysis of the overall characteristics of speech recognition,a recognition method combining HMM and improved neural network is proposed.The quasi-Newton algorithm is used to quickly converge in a short time and within the small number of network iterations.The advantage of this is trimming and optimizing the relevant parameters of the BP network to reduce the number of iteration steps of the network and to improve the overall performance of the network.After simulation test,it is found the proposed algorithm has certain advantages in speech recognition than traditional algorithms,and improves the speech recognition rate.Finally,the whole speech recognition process is simulated on the MATLAB7.0 platform.The characteristics of speech recognition rate under the same SNR and different signal and noise are analyzed respectively,and compared with the speech recognition rate of traditional recognition algorithms.The simulation results show that,based on Hidden Markov and improved neural network speech recognition,the number of iterations of model training is reduced,the training time of the model is saved,the recognition rate is improved,and the robustness of speech recognition is enhanced.