Research On Blind Source Separation With Low Complexity In Binaural Hearing Aid System

People with hearing loss usually have poor selective selection capability and find difficulty in determining target speaker,which requires that digital hearing aid can help distinguish the target and interference signal.BSS technology can cancel the interferences and at the same time preserve binaural cues through post-processing algorithms when used for speech enhancement in binaural hearing aids,which can help the users localize the speakers.This makes up for the shortcomings of the current adaptive noise reduction techniques.Therefore.BSS has great potential for application in binaural hearing aids.Convolutive mixing model is close to the actual environment and separation can be performed in time or frequency domain for convolutive mixture.The time domain methods suffer from slow convergence in echoic environment,while the frequency domain methods are computationally efficient,as the convolution in the time domain becomes simple element-wise multiplication in the frequency domain.Frequency domain independent component analysis(FDICA)is widely used.However,its high computational complexity fails to satisfy the requirement of low complexity for algorithms in binaural hearing aids.To solve the problem,frequency bin selection FDICA reduces the computational complexity by selecting limited frequency bins to perform ICA.The current frequency bin selection FDICA usually assumes the closely spaced microphones.There is no one aimed at the microphone spacing of dmic=0.15m,which is a typical distance between two microphones of binaural hearing aids.In other words,no frequency bin selection FDICA is suitable for the binaural hearing aids now.Therefore,the thesis does some research on the related algorithms and my contribution is as follows:1)A conventional FDICA algorithm is realized.As the key steps of FDICA,complex ICA and solutions for two ambiguities are introduced.The combined algorithm of FastICA and SNG ICA is used for ICA and the choice of nonlinear functions and parameters for the combined algorithm are discussed.Meanwhile,the permutation ambiguity is solved by using basis vectors and the scaling ambiguity is solved according to the minimum distortion principle.The extensive simulations show that the conventional FDICA has good separation performance.Finally,the computational complexity of each step of FDICA is estimated and the results show that ICA updates are the dominant cost.So limiting the number of the frequency bins to perform ICA can reduce complexity significantly and this is the basic idea of frequency bin selection FDICA.2)A frequency bin selection FDICA based on determinant is proposed for two microphones spaced dmic=0.15m in anechoic environment.The proposed algorithm adopts an innovative two-stage frequency bin selection procedure and a new separation algorithm for unselected frequency bins.The first-stage selection uses the determinant of the covariance matrix of the mixed signals as criteria to get the preliminary frequency bins.These bins are screened further by the second-stage selection based on boxplot for outlier detection.The left ones are the final selected frequency bins Two-stage selection guarantees good separation performance of the selected frequency bins.The rest are the unselected ones.The proposed algorithm uses the estimated relative mixing parameters to form separation matrices for the unselected frequency bins and solves the scaling problem at the same time.It is simple and effective.The extensive simulations show that when the threshold is set to 0.0 1 compared with conventional FDICA,the proposed algorithm reduces the running time by about 90%with a significant performance improvement,confirming that the proposed algorithm both reduces the complexity and improves separation performance.3)A frequency bin FDICA based on mutual information is proposed based on the one based on determinant.Only the first-stage selection criteria are modified to mutual information and the selection scheme is modified accordingly.The other parts stay unchanged.The two algorithms are compared by extensive simulations.With a small increase of the complexity due to estimations of the mutual information,the proposed one shows better optimal separation performance by adjusting a threshold.