The Study Of Dual-channel Speech Separation Technology For Smart Mobile Devices

In recent years,smart mobile communication devices have developed rapidly.Speech,as an important interface of communication,plays a dicisive role in people's lives.Sound source separation technology,which eliminates echo and environmental noise,plays a critical role.Blind source separation is one of the representative algorithms.A series of advanced algorithms for blind source separation come into being in complex application scenarios.In addition,using neural network as post filtering to eliminate residual noise is a common method.This thesis mainly studies the separation framework based on blind source separation algorithm and the post-filter technology based on recurrent neural network.The main work of this thesis is summarized as follows.First,the traditional separation algorithm calculates a large number of fast Fourier points in the iterative process,which leads to a large time delay and makes it difficult to meet the requirements of real-time response.To overcome this disadvantage,an improved scheme is proposed in this thesis,which sets up two filters to process simultaneously.The blind source separation algorithm in update filter calculates the weighting factor,and the output filter achieves the real-time processing and reduces the time delay using frequency-domain blocked method.Secondly,the speech after separation still contains residual noise.This thesis studies the signal processing and post filtering technology based on neural network.In order to meet the requirements of protecting speech quality and eliminating noise at the same time,the loss function based on speech protection is applied to recurrent neural network,and a new post filtering technique is proposed.The proposed algorithm can eliminate residual noise and protect speech quality effectively.Finally,the simulation results under different reverberation and different noise scenarios using real-world data show the effectiveness of the proposed algorithm.For the framework combining blind source separation and the delay block frequency-domain adaptive filter,it is found to not only improve separation effect but also reduce time delay effectively.As for residual noise,recurrent neural network post-filter technology based on new loss function can effectively extract desired signals and achieve better noise reduction results.