Deep Learning-based Speech Enhancement With Microphone Array

In modern communications,speech will inevitably receive external environmental interference,which is superimposed with clean speech in the time domain and frequency domain,resulting in the decline of communication quality.Speech enhancement is to restore the noisy speech as much as possible to the original clean speech.Microphone array speech enhancement can not only make use of the time-domain and frequency-domain information of speech,but also make use of the spatial domain information.This joint time-domain,frequency-domain and space-domain enhancement method can enhance the signal in the clean speech incident direction and suppress the signal in other incident directions,thus greatly enhancing the effect of speech enhancement.As a common algorithm in microphone array speech enhancement algorithm,GSC is simple and easy to implement.GSC can effectively suppress coherent noise,but can not effectively restrain the incoherent noise,especially the coherent and incoherent noise existing environment at the same time.Besides,the incoherent noise error of adaptive branch estimation is large.What's more,the spatial incoherent noise will be amplified when Subtracting the adaptive branch output from the fixed branch output.In view of the above shortcomings,we proposes a method named DNN-based Speech Enhancement with Microphone Array in this paper which can improve the accuracy of GSC noise estimation and make the GSC deal more effectively with the incoherent noise by adding a DNN-based post filter.We introduce the research background and significance of microphone array speech enhancement,and research progress at home and abroad and the related principles.Besides,the main work of this paper is as follows:1.Propose a new noise estimation method to solve the defects of large incoherent noise estimation errors of GSC,and prove it through experiments.Avoiding to amplify the the spatial incoherent noise,a model-based filter is used to take place of the subtraction operation between the fixed and adaptive branches in the conventional GSC,and then make an experiment to compare the performance of the model-based GSC algorithm with spectral subtraction,wiener filtering,MMSE and the model-based speech enhancement algorithm..2.In view of the shortcomings of model-based GSC algorithm,a method named DNN-based GSC algorithm is proposed in this paper,which uses the DNN to replace the GMM as a postfilter aiming at improving the effect of speech enhancement.Finally we make an experiment to compare the performance of the DNN-based GSC algorithm with GSC,DNN-based and the model-based GSC algorithm.The result of the experiment shows that the DNN-based GSC algorithm perform better than the contrast algorithm in different noise environments,and the speech quality is improved obviously.