Research And Implemention On Speech Separation Algorithm

Speech separation has been widely used in many practical tasks,e.g.daily life,ocean exploration,military interception.According to the number of mixed speech signals,there are two categories of speech separation methods including multi-channel speech separation and monaural speech separation.The methods applied to solve multi-channel speech separation include independent component analysis and kernel independent component analysis(KICA),while nonnegative matrix factorization(NMF)and deep neural networks are widely used to solve monaural speech separation problem.KICA applies kernel method to ICA to get higher accuracy,but the employed kernel matrix is quite big and results in high computational overheads.A usual solution to this problem is constructing low-rank approximation of the kernel matrix.Prior works make use of incomplete Cholesky decomposition(ICD)to construct such low-rank approximation.ICD does cut down the computational burdens of KICA.However,with the increase of the scale of dataset,both time and space complexities of ICD are unacceptable.This paper proposes a fast algorithm termed Nystr?m-KICA for optimizing KICA based on the Nystr?m method.In contrast to ICD,the Nystr?m method uses a much simpler sampling technique,and thus it can construct the low-rank approximation of the kernel matrix more efficiently.Inspired by the success of KICA,we found that Nystr?m method can be applied to accelerate kernel canonical correlation analysis(KCCA).We therefore further utilized the Nystr?m method to reduce the computational complexity of KCCA.In monaural speech separation,non-negative matrix factorization(NMF)is a popular method.However,traditional NMF cannot utilize the information in testing data.A semi-supervised method called transductive NMF(TNMF)is proposed to solve this problem.It has shown great power in monaural speech separation problem.In this paper,we implement TNMF-based monaural speech separation on many integrated core(MIC)architecture.The experimental results confirm the effectiveness of our work.