Research On Speech Dereverberation Based On MCLP Using NMF

As a kind of voice acquisition device,microphone array is widely used in the following aspects,such as teleconferencing,the conference rooms,and hearing aids and so on in a relatively closed acoustic environment.Because of the influence of reverberation and noise,it is difficult to obtain high quality speech signal.For this reason,a series of related field tests and researches on speech enhancement method in recent years have been carried out to deal with the undesirable factors such as reverberation and nosise.Moreover,as an important branch of speech enhancement,speech dereverberation always plays an important role in many areas such as speech recognition,sound source localization,seismic data processing and so on.So it is of great theoretical and practical significance to strengthen the research on indoor speech dereverberation.First,the generation of the reverberation,reverberation characterization and mathematical model of reverberation is briefly described.And then the main speech dereverberation methods are described in detail.Finally,according to the relevant theoretical knowledge,the research about speech dereverberation methods will be carried out in the actual enclosed acoustical environment.Studies have shown that the reverberation speech signal is usually divided into three parts: the direct speech signal,the early reflections and the late reverberation.While the early reflections can improve the speech intelligibility,the late reverberation mainly deteriorates the speech intelligibility due to overlap-masking.So the direct speech signal and the early reflections are usually choosed to as the desired signal to reconstruct.In the present study,the blind dereverberation method based on the multi-channel linear prediction(MCLP)mainly discussed and study in the short-time Fourier transform domain.And on the basis of this,some improvements are put forward to further improve the dereverberation performance.In the original method,the input signal is directly used to initialize the spectral variance of the target signal.In order to study the influence of covariance initialization,a covariance initialization scheme based on coherent-to-diffuse power ratio(CDR)is proposed.In addition,some researchs have shown that the short-time Fourier transform coefficient of clean speech has a certain sparse property and is more sparse than the short-time Fourier transform coefficient of the reverberation speech signal,thus through improving the sparsity of the output signal to yields a signal which better resembles a clean speech signal as the final target signal.In this paper,the best of the sparse property of clean speech signal and the decomposition property of NMF to will be fully used to apply it into the dereverberation method based on MCLP using a complex generalized gaussian prior(CGG).In the end,the final experimental results show that the improved methods can achieve a better dereverberation effect than the conventional dereverberation method based on MCLP under the same conditions.