| Abstrakt: |
Source separation by estimating the mixing matrix (A) in the underdetermined condition of Blind Source Separation (BSS) is playing a vital role where the number of sensors is less than the sources. Earlier methods are proposed based on Hierarchical Clustering (HC), Sparse Component Analysis (SCA), and Nonnegative Matrix Factorization (NMF) for mixing matrix estimation. In this work, these methods are verified in audio processing applications. From experimental results, it is observed that the estimated mixing matrix is significantly deviating from its actual mixing matrix. This causes degradation in the quality of separated audio sources. In this work, we proposed a new approach based on Successive Projection (SP) using convex geometry in the covariance domain. In the successive projection, frame-wise cross-correlations between the sources will be attenuated substantially. If exact locally dominant frames are not identified, then the SP gives degraded performance because in practice locally dominant frames are considered approximately. To overcome this, the Volume Minimization Criterion (VMC) is used to estimate the mixing matrix. Simulated results show the superiority of the proposed methods over HC, SCA, NMF based algorithms. |
Full text from SpringerLink