Research On Multi-speaker Recognition Technology Based On FMCW Rada

Speaker recognition has been used in finance and Internet.At present,most of the research in this area is based on microphone,which use a contactless way to collect signals.It has the advantage of low cost of use and high accuracy.However,the voice signal is easy to be copied,and it is easy to be attacked by criminals.Vocal vibration plays a key roll in pronunciation.Radar sensors have been used in the detection of vocal vibration.They have the advantages of high directivity.In this paper,we achieve speaker recognition based on the frequency modulated continuous wave radar.The main work and research contents are summarized as follows:(1)Introduced the principle of voice utterance,classic acoustic model and the working principle of FMCW radar to achieve voice detection,and introduced the working principle and advantages of multiple input multiple output(MIMO)radar.(2)A multi-speaker fundamental frequency extraction algorithm based on improved Variational Mode Decomposition(VMD)is studied.First,we estimate the direction of arrival of the received radar echo signal to locate the speaker.Then,we extract speaker’s signal through beamforming.Then,the phase fitting method is used to perform body motion compensation.Based on this signal,an improved VMD adaptive speech fundamental frequency extraction algorithm is proposed.This method uses the spectral envelope of the speech signal to obtain the initialization frequency and the number of modes of VMD.It can adaptively obtains the fundamental frequency signal component through energy judgment.Human body measurement experiments show the radar can effectively detect the vocal vibration,and the improved VMD can correctly extract the fundamental frequency.(3)A speaker recognition method based on speech fundamental frequency and characteristic parameters is proposed.First,we build a MLP-CNN model,then we study the speaker recognition method based on the fundamental frequency and the other model based on the fundamental frequency and voice features.Then we introduce the attention mechanism and enhance the learning ability of the model.The actual test results show that the recognition accuracy of the speaker recognition model is 82.8%.Verify the effectiveness of the method in this paper.