Research On Speech Enhancement Algorithms Of Circular Microphone Array

In our life,speech is one of the most important ways of communication,but the environment is full of various noises,which seriously affects the quality of speech.Therefore,speech enhancement technology is increasingly important.Single channel speech enhancement,as the most mature algorithm of speech enhancement,is difficult to obtain the ideal effect in the complex environment,while the microphone array can obtain the space-time information of the speech signal and has a high spatial resolution,so the microphone array is widely used in the field of speech enhancement.Because of the influence of the priori parameter estimation is not accurate due to the noise,the performance of the microphone array speech enhancement is still severely restricted.In order to improve the quality of speech,based on the circular microphone array,an improved minimum variance distortion response(MVDR)beamforming algorithm and its associated post filter system are proposed in this thesis.In the existing algorithms,the noise covariance matrix is updated by the adaptive method and the steering vector is estimated by the Generalized Cross-Correlation Phase Transform(GCC-PHAT)method in MVDR.The performance deteriorates rapidly in complex environments.For this reason,two improvements were proposed in this thesis.On one hand,the inversion of the sampling matrix in radar beamforming is used,and the noise covariance matrix is calculated using a dynamic smooth update method to ensure the stability of the noise covariance matrix estimation.Then the characteristics and practicability of dynamic diagonal loading technology and fixed diagonal loading technology are analyzed,the fixed diagonal loading technology is used to enhance the robustness of the system.On the other hand,based on the facts that the accuracy of the steering vector estimation is severely reduced in the complex environment and noise mainly pollutes formants,the Hilbert envelope of linear prediction residuals is used to estimate the time delay,which eliminates the error caused by the noise and improves the accuracy of the time delay estimation.Comparing the improved method with the traditional method,the simulation results shows that the improved time-delay estimation method can still maintain its performance at lower signal-to-noise ratios(SNR),at the same time,the improved beamforming system has greatly improved both in PESQ score and the word error rate(WER),even in the case of low SNR,the PESQ of improved beamforming system increased by about 1.0,and the WER dropped to about 1.42%.After beamforming,the coherent noise and non-coherent noise still remained in speech,so the Wiener post filter is introduced for further processing.Existing Wiener post filters are mostly assumed to be used in non-coherent fields,and the same weighting is used for the entire frequency band.Based on the fact that the actual noise field is a scattered noise field in our cases and the noise is highly coherent at low frequencies,a coherent function of the scattered noise field is used to estimate the power spectral density of expected signal and the noise signal,besides a dual parameter method is introduced to dynamically adjust the Wiener filter.Then dynamic updating is carried out in the low frequency band based on the input segmenting SNR,so that the Wiener filter can be updated adaptively according to the frequency and SNR,and the fixed Wiener filter is used in the high frequency band to suppress the residual noise with the lowest distortion possible.Comparing improved method with existing method,the simulation results shows that with the improved post Wiener filter,the low frequency noise is significantly reduced,meanwhile the speech distortion is not significantly increased,and the PESQ score is improved.A more robust parameter estimation method and a dual-parameter variable-Wiener filter suitable for the scattered noise field are adopted to the improved circular microphone array speech enhancement system,the single-channel and narrow-band signal processing method are modified to apply in the multichannel,broadband environment.At the same time,effective and economical microphone array is used in this system,and the system has good performance and strong stability under a variety of noise and SNR.Conclusively,this system is innovative and practical.