Research On Speaker Identification Based On Speech Processing

Face recognition,iris recognition,pupil recognition,fingerprint recognition and voiceprint recognition based on biometrics have continued to develop and made great progress with the development of science and technology.Because voice is difficult to imitate,and its data collection is convenient,simple and low cost.New technologies based on the characteristics of speaker voice signals to realize speaker identity recognition are emerging,which make the application of speaker recognition technology more and more,such as public security judicial verification,bank transaction system,mobile phone intelligent payment and voice control door and so on.Recently,the covid-19 pandemic has swept the world,and people have responded to experts' suggestions to wear masks to prevent infection,but it makes the system of identity recognition by face recognition in trouble;at this time,the advantages of speaker recognition are highlighted,so it is of great theoretical value and practical significance to carry out the research of speaker recognition based on speech processing.There are too many human factors in speaker recognition based on traditional feature parameter extraction and pattern matching,and the model can not fit the speaker characteristics when the amount of data increases.In this paper,a speaker recognition method based on convolutional neural network algorithm for training,learning and recognition of enhanced spectrogram is studied.The main work is as follows:(1)The research significance of speaker recognition is introduced,its development status is summarized,common speech feature extraction algorithms and traditional speaker recognition models are elaborated,and the advantages and disadvantages of different speaker recognition models are discussed.(2)Speaker recognition algorithms with different neural network structures are analyzed and discussed.In this paper,the principle of artificial neural network is introduced,and the speaker recognition based on depth neural network,delayed neural network and convolution neural network structure is discussed.At the same time,different activation functions and methods to prevent overfitting are discussed.The idea of speaker recognition based on convolution neural network is determined.(3)Combined with image recognition thoughts,the original speech of the speaker is converted into two-dimensional graph spectral information through preprocessing,and the spectrum image is enhanced by image enhancement technology.Several image enhancement algorithms are simulated on pycham platform,the advantages and disadvantages of each enhancement algorithm are analyzed,and the convolution neural network algorithm is proposed to train and learn the enhanced spectrum to identify the identity.(4)A speaker recognition system based on neural network is built.The neural network is based on lenet5 network.By comparing the influence of different learning rate,iteration times and parameters of network structure on the network recognition rate,the network structure and parameters for speaker recognition are determined.Then,the simulation test of the Chinese speech database of aishell is carried out by using this network model.The test shows that the algorithm is feasible.At the same time,QT creator speaker recognition application system based on pycham platform is designed.The generation and enhancement of speech spectrum and the result of speaker recognition are encapsulated to realize the visualization of speaker recognition in real time.