Speech Enhancement Method Fortibetan Speech Recognition In Lhasa Dialect

Speech enhancement is a method of noise reduction and speech enhancement which aimed at the problem of noise interference frequently encountered in voice transmission.In speech recognition,there will always have some noise in speech.Speech enhancement can effectively reduce background noise,enhance speech and improve speech recognition effect.This topic of thesis is originates the National Natural Science Foundation of China.It is necessary to solve the noise problem of speech recognition in the implementation of the project,and there is the purpose and significance of this study.The purpose of thesis is to solve the problem of noisy speech recognition and improve the quality of listening.The study is more effective in distinguishing noise and Tibetan Lhasa dialect enhancement system,and improving the accuracy of speech recognition.This thesis mainly discusses the technology of speech enhancement and the method of Tibetan speech processing.The following work has been carried out in the thesis.the technology of speech enhancement and the method of Tibetan speech processing are discussed and analyzed.The characteristics of Tibetan pronunciation and the related knowledge of Tibetan phonetics are summarized,and the knowledge of speech pretreatment and speech digital signal processing is briefly introduced.The standard method is used for the standard of speech enhancement language and the related process of acquisition.It introduces how to collect and make the experimental corpus,and how to make and test the test corpus.In the thesis,the traditional single channel speech enhancement method are selected.Several representative methods are selected in the traditional speech enhancement method.The principle of traditional speech enhancement algorithm is discussed.The block diagram of the speech enhancement system is given.The technical route of the traditional single channel voice enhancement and the work flow of the system are also introduced.The thesis disguss the result of experimental.At the same time,the thesis explores a new method of speech enhancement: The neural network is used for reference.The speech enhancement deep neural network model is innovatively proposed,and the model is used for noise reduction.The deepneural network model is adopted to train the noisy speech to complete the speech enhancement.The mathematical model of deep neural network is studied,and the principle of speech enhancement deep neural network is discussed.The speech enhancement depth neural network model of training is discussed,and the work flow of speech enhancement depth neural network is given: The feature of data nonlinearity is extracted,the training of depth neural network is adjusted,lock data feature and finish build model.The influence of the number of hidden layers on the enhancement effect of deep neural network is studied.The experimental results are obtained by the above two methods.The subjective and objective criteria are discussed and analyzed,and the results of speech enhancement are obtained by comparing the traditional single channel method and the deep neural network method.Adding the evaluation standard of speech recognition rate,which is used to judge the effect of the enhancement method.By analyzing the data of enhanced standard,the following conclusions are obtained.1)The deep neural network method is better than the traditional speech enhancement method on the subjective evaluation criterion.2)The deep neural network method is obviously superior to the traditional enhancement method on the objective standard.3)In the traditional single channel methods,the difference enhance effort gaps between the spectral subtraction method and the adaptive filtering method is not very large.The effect of the Wiener filtering method is better than the spectral subtraction.The approximate white noise is leaved after the logarithmic minimum mean square error method enhance and music noise is removed.The other three traditional methods have some music noise to stay.