Study On Blind Source Separation And Dereverberation Techniques For Multichannel Speech Enhancement

Sounds captured by distant microphones in a room often contain the direct sound,the reverberant part,and the interference from other sources.The reverberation and interference have greatly restricted the performance of audio and speech processing technologies in teleconferencing and far-filed human-machine interaction.As two kinds of speech processing techniques,blind source separation(BSS)can separate the target source while speech dereverberation can mitigate the unexpected late reverberation.Independent vector analysis(IVA)has been one of the popular frequency domain BSS algorithms,which finds statistically independent sources from mixture observations by utilizing higher-order statistical properties.However,its performance relies heavily on the proper modelling of acoustic sources.In order to achieve an interference-free separation,the distribution adopted in IVA should match the exact source distribution as closely as possible.Multichannel linear prediction(MCLP)is a blind dereverberation method based on the speech linear prediction model.In order to be applied in time-critical applications like telephony,the adaptive variant of MCLP method based on the weighted recursive least square(RLS)algorithm has been proposed.However,RLS algorithm is prone to numerical instability and the value of the forgetting factor affects its convergence speed and tracking ability.Thus,it is significant to design an adaptive mechanism for the MCLP algorithm with stable and strong tracking ability.Moreover,the separation performance of conventional BSS methods deteriorates as the reverberation time increases,while most existing dereverberation methods rely on the assumption that there is single source in a room.Therefore,a multichannel speech enhancement system combining blind source separation and dereverberation should be designed to enhance the separation performance in the high reverberant environment.In this thesis,both the IVA algorithm and MCLP algorithm have been studied.The main research content and innovation of this thesis are summarized as follows.(1)The complex generalized Gaussian mixture distribution with weighted variance is introduced to IVA algorithm,which is capable of capturing both non-Gaussianity and non-stationarity.The majorization minimization framework is adopted for the IVA optimization.The experimental results reveal that the proposed algorithm attains the best performance when the shape parameter is within the range of 1.6 and 2.Compared with the conventional IVA methods,it can improve the signal-to-interference ratio of 2?5d B and the signal-to-distortion ratio of 1?3d B.(2)An adaptive MCLP speech dereverberation method based on QR decomposition recursive least squares is proposed,and a time-varying forgetting factor control mechanism is designed to adapt to dynamic acoustic scenes.Experimental results show that the proposed algorithm can improve the signal-to-noise ratio by 2d B,and enjoy fast tracking capability and numerical robustness.(3)Two kinds of multichannel speech enhancement methods combining blind source separation and dereverberation are designed.Experimental results show that the MCLP method can effectively suppress the late reverberation components of the mixed speech signals.It can be concluded that using MCLP algorithm as the preprocessing of the separation algorithm is beneficial to improve the speech quality.