Study On Speech Enhancement And Post-Filtering Technology Based On Microphone Array

In recent years,thanks to the data rate and bandwidth advantage of 5G technology,speech technology has gained in-depth development.Speech interaction technology has become a key research target for current researchers and engineers.Microphone array is the front-end enhancement technology in the speech interaction process.Compared with single-channel speech enhancement,it can obtain the spatial information of the signal so as to achieve a better suppression effect on the noise in all directions.This thesis mainly studies the microphone array speech enhancement and its post-filtering technology,focusing on the analysis of the algorithm performance of the beamformer,and proposes a new frequency-domain adaptive blocking matrix on this basis.In addition,in view of the poor suppression of incoherent noise by the beamformer,this thesis also proposes an improved post-filter,which can further reduce the incoherent noise in the output signal of the beamformer.First of all,thanks to the spatial information obtained by microphone array,it can estimate the time delay and perform beamforming in a certain direction,while suppressing noise in other directions.Although this method can effectively suppress noise,there are still many improvements to be expected.The speech leakage problem of the blocking matrix has not been perfectly solved.In response to this problem,this thesis proposes a new frequency-domain adaptive blocking matrix design for generalized sidelobe cancellation.This method can effectively reduce speech leakage and at the same time reduce the false cancellation of reference noise signals,thereby reducing speech distortion.Secondly,although the microphone array can achieve better noise suppression than the single-channel processing method,there still exists massive incoherent noise.To conquer this drawback,this thesis puts forward an enhanced post-filter,which can be realized using a simple way and has low calculation complexity.Compared with other post-filtering methods,it can better eliminate incoherent noise and significantly decrease the voice distortion.Finally,by conducting the various experiments,we demonstrate the effectiveness of the two proposed methods.The improved beamforming algorithm enhances the noise suppression effect on low-frequency signals,and at the same time,the suppression ability of reverberation is also enhanced,which is equivalent to indirectly enhancing the robustness in a complicated condition.The enhanced post-filtering is also better than the conventional baseline method in processing incoherent noise and can significantly decrease voice distortion.The combination of the two methods has enhanced the overall capability of the system.