Audio Capturing And Enhancement Based On Microphone Arrays

In typical applications of speech capturing and communication,the desired signal is always corrupted by different kind of interferences,including background noise,competing speech,reverberation,and far-end/playback echo signals.Compared with a single microphone,microphone arrays are capable of utilizing acoustic spatial information,thus extracting speech more effectively from target direction and suppressing noise better.Therefore,they are widely used in hands-free communication,hearing aid,human-computer interaction and etc..In this thesis,audio/speech capturing and enhancement technologies based on microphone array are investigated,with primary attention on the beamforming,array element distribution optimization and post-processing algorithms.Some commonly used microphone array models and beamforming methods are introduced,along with some evaluation metrics of beamforming algorithms.Based on the beamforming methods,a procedure for designing microphone array used for high quality audio recording is proposed.Because the high frequency details need to be preserved in such an application,the intervals of the adjacent microphones must be small enough to meet the Nyquist sampling theorem,possibly leading to impractically large amount of microphones.By optimizing the array element distribution,the number of microphones is reduced to an acceptable range for real applications.Experiments are carried out to demonstrate that the microphone array designed with the proposed method outperforms the same-length shotgun microphone significantly in directivity while maintaining a comparable low self-noise level and good frequency responses.For a microphone array with a small size and limited number of elements,beamforming alone often cannot achieve satisfactory speech enhancement.In order to further improve the performance,a deep learning model is proposed for post-processing after beamforming.By generating the training set in various acoustical environments and sharing weight between different channels,the model generalizes quite well to the experimental data collected by a microphone array.The model is tested using speech objective evaluation metrics and speech recognition systems,validating its superiority over the traditional signal post-processing method.Because the model is simple and generalizes quite well,it has a potential to serve for array systems in real application.Finally,an overview of the recent development of deep learning based speech enhancement/separation is presented.Several cutting edge models are introduced and the possible development in the future is discussed.