Research On Voiceprint Assisted Security System Based On Deep Learning

In recent years,with the rapid development of China's economy,the level of economic development in various regions appears certain imbalance,and the scale of population flow increases accordingly,which brings certain unstable factors to social security.However,the traditional methods based on video monitoring and target detection require a transmission bandwidth of several to tens of megabytes,and the storage demand is also very high.Therefore,it is particularly important to develop an effective and intelligent security assistant tool,which can reduce the dependence on high-definition video.Voiceprint recognition is to input the audio information into the deep learning model and extract the texture features from the target voice.In contrast,other types of biometric recognition are easy to be affected by various interference factors,while voiceprint recognition is less affected after effective model processing.The cost of equipment needed in the identification process is also low,which is an effective auxiliary means under the current video security.In this thesis,a voiceprint assisted security system based on deep learning is studied,which mainly includes two aspects: hardware acquisition unit and deep learning model design.On the basis of the existing research,the hardware circuit design of voice acquisition is improved,the function of detecting the start and end points of voice signals is added,and the effective deep learning algorithm is combined to optimize the structure of voiceprint recognition system and improve the simplicity and accuracy of voiceprint recognition system.The main contents of this paper are as follows:1.This thesis introduces the industry background and monitoring necessity of monitoring field at home and abroad,analyzes the current situation of voiceprint recognition technology at home and abroad,and expounds the theoretical basis of voiceprint recognition;analyzes the requirements of voiceprint recognition system,including the business and functional requirements of security voiceprint technology,and further expounds the feasibility of auxiliary security system based on deep learning voiceprint recognition on the basis of demand analysis2.Optimize the existing voice acquisition circuit.The optimized hardware circuit of speech acquisition includes: pre-processing the acquired signal with microphone audio signal amplification module;signal conditioning with input buffer circuit designed by ad8656 chip;analog-to-digital conversion with cs5341 chip for the conditioned speech signal;receiving the speech data transmitted by ADC chip through SPI interface of STM32,and transmitting the data through USART peripheral Send to Wi Fi module,and finally transmit voice data to cloud as soon as possible through cc3100 wi fi module.3.The existing voiceprint recognition system has some problems,such as low anti-interference ability,weak noise reduction ability,unstable and incomplete feature extraction of voiceprint by MFCC(Mel frequency cepstrum coefficient).In view of these problems,this paper adopts the following four ways to improve and optimize:(1)Using "thin resnet-34"(3 million parameters)as the backbone architecture of the model to extract frame level features,the model greatly reduces the number of parameters,shortens the training time,and reduces the energy consumption;(2)The dictionary based netvlad is used to aggregate the features of different time for end-to-end training;(3)The softmax function was optimized,and margin softmax(am softmax)was used;(4)The overall performance of the system is closely related to the quality of speech preprocessing.Therefore,the end-to-end learning method is used to extract the effective segments in the speech,and combined with spectrum gated noise reduction,improve the quality of speech preprocessing.Through the experimental analysis,the model designed in this thesis is compared with other systems(i-vectors and TDNN(x-vectors)).In pure speech environment,EER(error rejection rate)of the model in this paper decreased by 1.37% and 0.57% respectively.In noise environment,EER decreased by 2.32% and 1.02% respectively.The experimental results show that the voiceprint recognition model based on thin Res Net-34 can improve the recognition ability and anti-interference ability of the system.The voiceprint assisted security system based on deep learning makes up for the shortage that the current mainstream video monitoring can not be stored for a long time.For the emergency audio data can be sent to the cloud server through the network,the bandwidth demand is low and the energy consumption is less,so it has a wide range of application prospects.