Study On Farfield Microphone Array Speech Enhancement Technology

In recent years,with the rapid development of artificial intelligence,speech signal has received extensive attention as an important interface for human-computer interaction,and research on speech technology all over the world has set off a new climax.The microphone array utilizes the spatial information of the speech signal,which explore more space-time information than the single-channel method,has the advantages of higher gain,beam flexibility,and strong noise suppression.The Generalized Sidelobe Cancelling(GSC)algorithm transforms the constrained problem in beamforming into an unconstrained problem and does not require a priori information for noise estimation.It is widely used in engineering.However,in practical application scenarios,due to problems such as microphone mismatch,inaccurate delay estimation,expected direction error,and reverberation,there will be leakage of the desired speech signal in the blocking matrix,which will lead to the problem of desired signal cancellation in the adaptive noise cancellation module.In addition,in the process of speech enhancement,it will cause distortion of some speech features,reduce the noise robustness and recognition rate of the speech recognition system,so it is necessary to enhance the feature of the speech signal.This thesis mainly studies robust GSC adaptive beamforming and speech feature enhancement algorithms.The main research contents are as follows:Firstly,the conventional GSC algorithm is prone to the desired signal cancellation during adaptive noise cancellation,and the conventional fixed beamforming algorithm is used in the upper branch to make the GSC output signal less robust.This thesis proposed an improved robust GSC algorithm.In the adaptive noise cancellation module,the coherence and energy ratio of the signal are used to jointly control the update of the adaptive noise canceller coefficients,and the robust superdirective beamforming algorithm is used in upper branch.The proposed improved algorithm not only effectively reduces the cancellation of the desired signal,but also further improves the robustness and low-frequency characteristics.Secondly,for the problem of distortion of speech features due to noise in far-field microphone array speech recognition,this paper deeply analyzes the signal feature domain and speech feature enhancement based on signal processing,and integrates multi-channel speech presence information into Wiener filter speech feature enhancement.Compared with the traditional Wiener filtering feature enhancement algorithm,the improved algorithm can further suppress the residual noise of the GSC output signal,thereby improving the noise robustness and recognition rate of the speech recognition system.Finally,the simulation of multi-channel speech data recorded in real scenes shows that the robust GSC adaptive beamforming algorithm based on coherence and energy ratio proposed in this thesis,and the feature enhancement algorithm based on the existence of multi-channel speech,can both effectively maintain the desired signal undistortion,while improving the noise robustness and recognition rate of the recognition system.