Research On Microphone Array Speech Enhancement Algorithm In Multiple Speakers Scenario

With the popularization of intelligent products in recent years,the realization of human-machine interaction through speech and other methods has received widespread application and attention,and smart home products represented by smart speakers have played an important role in people’s lives.As the application scenarios of smart products become more and more diverse,the environmental of speech interactions is becoming more and more complicated,especially in multi-speaker scenarios such as indoor homes,conference rooms and offices,the presence of interference speaker affects the voice quality of the target speaker,and reduces the voice interaction experience of the target speaker.The purpose of speech enhancement is to extract relatively pure target speech signals from disturbed speech,which is one of the key technologies of human-machine speech interaction.Microphone array composed of multiple microphones in accordance with a certain geometric shape can form a beam in a desired direction,enhance the target speech signal while suppressing interference signals from other directions,and can effectively improve the voice quality compared to a single microphone.It is widely used in audio and video conferencing products,smart homes and smart vehicle systems.The main research directions of microphone arrays are sound localization estimation and speech enhancement.Aiming at the problem of multi-speaker scene in indoor human-machine interaction,which affects the quality of the target voice signal due to the interference of the speaker,this thesis proposes a new multi-source positioning algorithm and beamforming algorithm,which is verified the performance of the proposed algorithm by experiments.The main work of this thesis is as follows:1.Aiming at the problem of the decrease of positioning accuracy of common sound localization algorithms such as delay and sum algorithm,steered response power with the phase transform algorithm,time difference of arrival algorithm and compressed sensing-orthogonal matching pursuit algorithm in multi-speaker scenarios,a distributed compressed sensing-simultaneous orthogonal matching pursuit algorithm is proposed by transforming the problem of multi-speaker localization into the problem of joint sparse reconstruction of signals.Simulation and experimental results show that the proposed algorithm improves the ability of sound localization in multi-speaker scenarios.2.Based on deep learning and combined with mixed training,the spatial directivity function in the form of Fourier logarithmic amplitude spectrum of the desired signal is taken as the optimization objective to optimize the spatial directivity of the desired direction of the microphone array.Experiments show that the proposed method is effective in improving the beamforming ability of microphone arrays in multi-speaker scenarios.