Research On Sound Source Localization Algorithm Based On Neural Network

As one of the important ways of expressing and receiving message in people’s communication,sound contains a lot of useful information,the usage of sound and application scope covers all aspects of life.Nowadays,with the continuous improvement of artificial intelligence technology,it is more and more common to use voice information as the input of various applications.At the same time,many systems and scenarios not only have demand for sound content,but also analyze the location of the sound source.At present,many practical results have been achieved in the research and application of sound source localization,but there are still many specific problems that need to be solved urgently.For example,under the interference of reverberation environment and external noise,the performance of the system for sound source localization will be reduced to varying degrees.Therefore,the in-depth study of sound source localization still has broad practical significance.This thesis first analyzes the real situation of sound source propagation in the room,that is,the influence of room impulse response(RIR)on sound source localization.An improved room impulse response simulation model is proposed and compared with the Single-and Multi-channel Audio Recording Database(SMARD)to obtain the simulated localization acoustic environment.Then,the theory of traditional microphone array sound source localization technology is studied,and three traditional sound source localization algorithms are mainly analyzed,namely:Steered beamforming,Spectrum estimation with high resolution and Time difference of arrival.Considering the geometric position relationship of the microphone array and the characteristics of the beam formed by the weighting of the received signal of the microphone,an improved localization algorithm based on the combination of PHAase Transform(PHAT)and Smooth COherent Transform(SCOT)is proposed.Finally,we combine neural network and sound source localization technology.According to the characteristics of neural network and sound signal,Res Net is used as the neural network model.Combined with the improved RIR simulation model,the sound signal received by the microphone array is calculated with GCC-PHAT-SCOT and the output of the calculation is used as the input feature of Res Net for training,and finally the result of sound source localization in the room is obtained in the test set.The experimental results show that the sound source localization algorithm based on neural network has high localization accuracy under different reverberation conditions and signal-to-noise ratio environments.