Research On Microphone Array Sound Source Localization And Beamforming For Speech Interaction

In recent years,the human-computer interaction mode with":intelligent speech recognition"as the core is in the golden period of its rapid development,and its momentum is still climbing.In addition to the excellent performance of the speech recognition model,the intelligent speech system is closely related to the microphone array technology used to collect speech signals.By collecting and processing the spatial characteristics of the target sound source signal,the microphone array can obtain the desired signal with high quality and play the role of sound source location and speech enhancement in the system.With the continuous improvement of people's demand for material life and quality,many products such as "intelligent audio" have come out and entered each family,providing convenience for human-computer interaction.The microphone array technology,which plays a key role in this process,mainly includes sound source positioning and beamforming.Through these two contents,high-quality voice recognition and subsequent feedback can be acquired for smart home products.In this paper,a deep learning based method for sound source location and beamforming in microphone array technology is proposed under the daily indoor"intelligent sound"working environment,and experimental verification is carried out.The specific work is as follows:(1)The conventional sound source localization techniques,such as DS,TDOA and SRP-PHAT,are introduced respectively,and the advantages and disadvantages of each algorithm are analyzed in terms of sampling rate,directivity,complexity and practical scene application,and puts forward will be based on the compression perception of microphone array localization algorithm is applied to the source environment,this method on the basis of the CS-OMP algorithm,delay relationship between arrays by using mixed matrix on the room impulse response of structure directly,in order to realize the sound source localization,and compared with the traditional method to complete the performance of the validation.(2)The technology of deep learning is introduced into the training of beamforming,and an independent neural network is constructed for the training of microphone array multi-channel beamforming.Based on the Tensorflow platform,a deep learning neural network that can be used for beamforming is constructed from scratch,including preprocessing module,neural network module and signal reconstruction module.Using average energy beamgraph and WER,the beamforming network is compared with the traditional linear beamforming algorithm.Experimental results demonstrate the feasibility and effectiveness of introducing deep learning into beamforming.