Design Of Voice Print Recognition System Based On RBM

Voice print recognition was one of the biometrics technology.Nowadays,data security was getting more and more attention.Compared with traditional recognition methods,voice print recognition had many advantages,such as difficulty in forgery,low cost,easy acceptance,easy voice acquisition,and support for remote authentication.Because of that,it had been widely used in finance,public security,intelligent building and other fields.At present,most voice print recognition models cannot solve the problem of Chinese speech recognition,because they were shallow network structures which were lacking the ability to describe voice print features.In the practical application process,the poor robustness of the model was easily disturbed by the external noise,which resulted in great deviation in the recognition results.Therefore,the study of voice print recognition system was of great practical significance.In this paper,the open source Chinese speech database was taken as the research object,and the Restricted Boltzmann Machine theory was applied to the study of voice print recognition system.It improved the ability of feature parameters to represent Chinese speech,built a deeper network structure model.In this way,the anti-interference ability of the model was improved,and finally the voice print recognition system was suitable for Chinese speech.This paper was developed on MATLAB R2017 a platform,algorithmically improved the traditional voice print recognition system.A voice check-in system based on unsupervised learning method was established by combining hard and soft methods.The main research contents were as follows:(1)The feature extraction method of speech signal was improved.The pre-processed speech signal was initially processed by using the MEL frequency cepstrum coefficient,and then the extracted vector was dimensionalized by using the restricted boltzmann machine network.By reconstructing the error curve,the structural parameters of thenetwork were adjusted to improve the representational ability of feature parameters after dimension reduction.(2)The network structure of voice print recognition model was improved and the structure frame of deep confidence network hidden markov hybrid model was constructed.From the perspective of improving the utilization rate of label missing data,the deep confidence network was used for unsupervised learning of speech after feature extraction to obtain more accurate observation rate of sample data.Then the probability prediction of speech category was completed by hidden markov model,and the final result of voice print recognition was obtained after comparison.Through design comparison experiment,it was found that the improved model recognition rate was higher after using depth confidence network.After optimizing the structure parameters of the network by reconstructing the error curve,it was found that the recognition rate was higher when the network was made up of 3-layer restricted boltzmann machines.(3)The hardware platform of voice check-in system was built.The DE2 development was boarded as the main platform for hardware design,Quartus II software compiled it.Through the WM8731 chip on the development board,the collected voice was converted from modulo to digital.The converted data were stored by the IS61LV25616AL-10 chip for subsequent computer software processing.S7-200 CN was used to constitute the controller of the system,and V4.0 STEP 7 software compiled the control part.Combined with the expansion module of the controller itself,the integrated control of voice input terminal,gate switch,electric lock,door resistance,card reader and other modules was completed.Some ports were reserved for the subsequent function expansion.The hardware part added additional relay control switch circuit,which can ensure the company,school and other places accessed control system in the extreme environment of normal operation,realized the voice pattern recognition system digitization and intelligence.