| Non-negative matrix decomposition (Non a negative Matrix Factorization,)referred NMF method, which is a new dimension reduction method in dealing withthe variety of data is a very favorable way, this method has separated data partialresponse characteristics of things, there is a wide range of applications, which askedfor the NMF method and its improved algorithm SNMF algorithm in monauralspeech separation, the main contents are the following:1. analyzes and summarizes a number of issues and theoretical system ofseparation of voice NMF methodNMF research status on leads and abroad for a simple overview of the currentnumber of systems and gives voice separation, gives the basic steps to separate monoaudio.2. the basic definition of NMF to make an explanation based on this proposedNMF objective function related to two types, one is the Euclidean distance function,one is KL divergence objective function, combined with the two iteration rule kind ofobjective function gives the basic derivation method, and finally prove theconvergence of these two objective functions, decomposition results prove theexistence made.3. shows three different methods at NMF method initialization, the initializationmethod of PCA, an improved method is to initialize the PCA SPCA, the last one isthe FCM.4. describes the basic features of speech, our hearts go here from the physicalunderstanding of the basic processes of human voice separation, thereby abstractingthe basic use of computers separation of voice processing steps. Finally, the basicblock diagram of speech separation.5. combined with the basic NMF algorithm and improved algorithm SNMFmonaural speech separation algorithm gives comparative results of the experimentprocess, we compare two ways one is accurate comparative rate of source separation,one is the same number of separate sources the time spent compared to the conclusionthat the first aspect SNMF stronger than the basic NMF algorithm, but two in thesecond aspect largely unchanged. |
PDF Full Text Request