Research On Application Technology Of Voiceprint Recognition System Based On Residual Network

With the development of biometrics technology and the improvement of people's pursuit of quality of life,Voiceprint Recognition(VR)is gradually applied to all aspects of life,for example,the application of VR in artificial intelligence.In addition,with the global spread of the Novel coronavirus pneumonia,VR has become a research hotspot due to its non-contact advantages.People's requirements for the real-time,accuracy,and application ecology of VR algorithms are also higher and higher.The VR algorithm based on Res Net has good real-time performance and high recognition accuracy.Therefore,research on VR system based on Res Net is of great significance.VR often collects noise in the voice collection stage,and the existing Voice Activity Detection(VAD)algorithms for VR based on Res Net are Ineffective.In view of the above problems,this thesis improves the speech processing flow,analyzes the requirements of the VR system based on Res Net,designs and tests the function of the system.The main work of this thesis includes the following aspects:Firstly,The theory and method of VR based on Res Net are studied,including data collection,preprocessing,feature extraction,training model and recognition.Aiming at the problem of Category score decreased of speech with noise,an adaptive decision threshold adjustment method is proposed,and starting with system requirements,software and hardware scheme of VR system based on Res Net is designed,appropriate hardware is selected and corresponding software environment is built.Secondly,the software and hardware implementation of VR system on Res Net was completed.Software has three parts,including: 1)The VR algorithm;2)VR system software on PC;3)VR client software on ARM.The system can recognize voices collected in real time on PC or ARM.Aiming at the two problems of large influence of noise on recognition results and poor performance of endpoint detection,the LMS speech enhancement module was added in the pre-processing and the dual-threshold method was used to implement the VAD module.the system hardware platform was built,On the JZ2440,voice acquisition module was completed,and send the voices to PC,and show the result from PC.Finally,the algorithm performance and function on system are tested.It shows : 1)For the Res Net to replace the convolution kernel,the training speed can be improved without reducing the recognition accuracy;2)the recognition accuracy of the VR algorithm based on the Res Net has been improved after improving the speech processing flow and adding an adaptive decision threshold adjustment method.The system combines technologies such as LMS,speech energy spectrum,network transmission,and Res Net to achieve various functions.It has some value for the application of the VR based on Res Net.