The Research And Implementation Of Parallelization Method For Mixed Sound Signal Identification

The sound that people hear is often made up of multiple sounds,from the mixed sound signals to quickly and accurately distinguish the people’s interested ones has always being the hot research.Although the traditional method can make simple sound source discrimination,when it comes to the large amount of complex sound signal processing,the real-time and accuracy of application is limited.However,with the advent of the era of artificial intelligence,new technologies,such as deep learning and GPU parallel computing,provide a solution for the large amount of sound signal processing.To this end,this paper designed a parallelization method for mixed sound signal discrimination,and has done the following work:1.This paper analyzes the domestic and foreign research status and development trend of mixed sound signals,in this context,mixed sound signals are as the research object,the relevant knowledge of mixed sound signal discrimination and GPU parallelization is introduced,the common methods of the separation of mixed sound signals and sound source discrimination are studied.2.Preprocessing is carried out for mixed sound signals,including de-averaging,whitening and so on.The fast-ICA algorithm based on negative entropy is selected to separate the mixed sound signals,the reason for the restriction of rapid discrimination in the separation process of mixed sound signals is analyzed,and accelerated improvement is performed by GPU parallelization.3.Multiple eigenvalues are extracted for separated sound signals,and the extracted eigenvalues are fused to consist of complex eigenvalues,then the sound source discrimination is performed.In the process of discrimination,due to the lack of learning ability of the traditional neural network,the deep neural network is introduced to design the sound source discrimination model based on deep belief network,which improves the accuracy of the mixed sound signal discrimination.4.Because of large amount of sound signals need to be processed,and sound signals have the characteristics of consistent method and strong independence in the process,GPU parallelization method is used to respectively optimize the Fast ICA algorithm based on negative entropy,eigenvalues extraction and training process of deep belief network model,the processing efficiency of the mixed sound signal discrimination is improved by using this method.Through the simulation and experimental verification,the efficiency of the separation and discrimination of the mixed sound signals is improved by using the method of GPU parallelization,and the requirement of real-time is met.At the same time,using complex eigenvalues based on multi eigenvalues fusion as the input data and the sound source discrimination model based on deep belief network can improve the accuracy of mixed sound signal discrimination.