Research On Key Technologies Of Audio Signal Separation In The Case Of Multiple Sound Sources

Nowadays,smart voice-UI devices are entering ordinary families naturally,meanwhile smart devices such as smart speakers and smart appliances have appeared in people's daily lives.At the same time,smart mobile devices have gradually begun to be equipped with more voice functions.Voice wake-up,voice-print and other functions are being integrated into mobile phones and tablets.However,how to use the microphone array board configured in smart devices to solve the problem of noise reduction in far-field voice pick-up in a noisy scene is still a major problem to be solved in both academia and industry.At present,the most important reason affecting the use of current intelligent voice devices is the complex noise scenarios.Indoor reverberation,indoor noise,competing speakers,etc.are all important factors that make up a daily noise scene.For the above noise scenarios,research on noise suppression algorithms with excellent performance has become the direction that academics and speech practitioners work together.For different types of noise scenarios,we usually choose different noise suppression algorithms.The current mainstream noise suppression algorithms include noise suppression,echo cancellation,beamforming,and so on.The noise suppression algorithm is to strip the clean speech signal from the mixed signal with noise collected by the microphone array.This paper is mainly about the research of audio signal separation algorithm from noise suppression algorithm.It focuses on the audio signal separation algorithm based on deep learning algorithm and microphone array,and proposes several optimization methods for the performance and algorithm architecture of audio signal separation algorithm.Audio signal separation,that is,the process of extracting the target speech from the collected mixed speech mixed with the interference signal.Among them,the collected interference signals may be Gaussian noise,music noise,or other competitive human speakers.This article at first introduces the principles and basic architecture of single-channel audio signal separation algorithm based on deep learning and multi-channel blind source separation based on microphone array.After that,based on the deep learning single-channel audio signal separation algorithm,this paper proposes two phase compensation training targets.The phase compensation separation method is used to optimize the separation performance of the channel audio signal separation algorithm.For the multi-channel blind source separation algorithm based on the microphone array,this paper uses the traditional sound source localization algorithm to optimize the initial value of the blind source separation algorithm for independent vector analysis,thereby saving the separation performance of the independent vector analysis algorithm and reducing the iteration number of cycles.Finally,we focus on the under-determined microphone array audio signal separation algorithm.We propose an under-determined multi-channel blind source separation algorithm that uses a combination of single channel audio signal separation algorithm based on deep learning and a multi-channel blind source separation algorithm based on microphone array.The separation performance has been experimentally verified.Finally,by comparing speech quality indicators along with signal-to-noise ratio indicators,we prove the effectiveness of our proposed separation method.