Study On Speech Enhancement Based On NMF Algorithm

Speech enhancement is an important branch of speech signal processing,to improve all kinds of noise speech quality and intelligibility.It is also an important pre-pretreatment technology of speech communication,speech recognition,and speech compression coding and digital processing system.With the limitation of conventional speech enhancement algorithm to handle non-stationary background noise and thus some new algorithms have been put forward in recent years.Speech enhancement algorithm based on nonnegative matrix factorization(NMF)is one of the most effective and prominent methods,which is a machine learning to analyze the desired data from a large number of voice data is one of current researching hotspots of speech signal processing.Based on above background,the main work is to analyze and summarize the speech enhancement algorithm under single channel system and multichannel system and its new variants is proposed.The based-NMF speech enhancement algorithm and its variants are proposed on rule of supervised learning principle.Firstly,the limitation of conventional speech enhancement algorithm is analyzed and studied.The basic principles and characteristics of the nonnegative matrix factorization are described,and the convergence and initialization of this algorithm are studied.Secondly,one main problem of the supervised learning algorithm is the existence of a mismatch between the characteristics of the training and test data,and the paper present the nonnegative matrix factorization algorithm is to add explicit exact regularization terms to the NMF cost function,such as temporal continuity or statistical priors of the magnitude spectra and sparseness which in turn leads to improve the accuracy of the collection of voice data by strengthening the division and discrimination of different dictionaries.However,the conventional speech enhancement system is based on the short-time analysis modification synthesis(AMS)framework.In most cases,it is considered that accurate spectral amplitude estimation plays a more important role than improving the phase spectrum in the sense of the perceived signal quality,so the influence of phase spectrum distortion is ignored.Based on this fact,this paper presents the basis compensation algorithm for NMF-based speech enhancement through the modification of phase.Based on the mathematical statistics analyze,we analyze and study the algorithm of nonnegative matrix factorization,and improve the speech quality by using the method of speech existence probability and phase spectrum modification.Finally,in order to overcome the conventional multi-channel nonnegative matrix factorization algorithm is not only easy to fall into local optimal,but also higher calculation complexity,which limits its application.Therefore,a new multi-channel nonnegative matrix factorization model is proposed.The model is mainly carried out by modeling the spatial covariance of observation data,combined with the property of matrix trace cleverly introduced the cost function,and then by using the principle of supervised learning algorithm,the mixed signals picked up by microphones are analyzed and studied to achieve faster and better noise reduction performance.The experimental results show that the proposed algorithm can not only reduce or suppress the background noise,but also improve the convergence speed,which makes the algorithm more practical.