Research On Sound Source Localization Based On Microphone Array And Neural Network

In recent years,with the development of smart products,sound source localization technology has been widely applied to intelligent robot,noise detection,honking capture,intelligent home and other scenes,and has become a research hotspot in the field of acoustics and signal processing.In practical application scenarios,sound propagation will be affected by reverberation effect.Both the traditional sound source localization technology and the deep learning-based sound source localization technology,the localization accuracy will be seriously decreased in complex environments.Therefore,it is very important to enhance the precision of sound source localization under the background of noise and reverberation.On the basis of the traditional sound source localization methods,we introduce the technology of microphone array and neural network model to investigate and study the sound source localization performance in complicated acoustic environment.Based on the microphone array signal model,indoor reverberation model and room impulse response model,this thesis introduces the fundamental of sound source localization,which provides a theoretical basis for the subsequent research.The main work and contributions of the thesis are as follows:(1)It becomes a challenging task to accurately determine the location of sound source using time delay estimation techniques under the background of noise and reverberation.In order to improve the localization accuracy,this thesis utilizes the good learning characteristics of back propagation neural network.The localization algorithm combining time delay estimation with BP neural networks is studied,and the localization algorithm based on cross-correlation sequence and BP network is proposed.Firstly,the localization algorithm based on GCC-PHAT delay estimation is introduced for double five-element cross array.Then,a localization algorithm based on GCC-PHAT delay estimation and BP network is proposed.According to the parabolic fitting fractional time delay estimation has higher accuracy,a localization algorithm based on parabolic cross-correlation time delay estimation and BP network is presented.Finally,by analyzing the main factors of influencing the time delay estimation,a localization algorithm based on cross-correlation sequence and BP network is proposed.Simulation results indicate that all proposed algorithms have better localization effect in comparison with traditional localization algorithm based on GCCPHAT time delay estimation,and each of the latter algorithm is more accurate than the former one.Moreover,the localization algorithm based on cross-correlation sequence and BP network also has good localization effect under the condition of low signal-to-noise ratio and high reverberation.(2)At present,sound source localization methods based on common array and deep learning have achieved good localization effect,but common stereo microphone array is not easy to be placed,dual microarrays are easy to be placed,and there are few researches on sound source localization methods based on dual microarrays and deep learning.Therefore,a sound source localization method based on dual microarrays and convolutional neural network is proposed.In this method,the maximum GCC-PHAT position information and the maximum possible delayed sampling point sequence values are extracted as the input of the network,and provides the sound source’s 3D position coordinates for the output.The network model for sound source localization is obtained through training.Experimental results show that the proposed sound source localization method achieves higher localization performance and capability of robustness under various acoustic environments.In order to further improve the localization performance based on dual microarrays and convolutional neural network model,the performance of different input characteristics is studied.