Research Of Generalized Gaussian Distribution Implementation In Music Blind Separation

Music signal is well known as a random signal with its probability density function (PDF) describing the statistic characteristics comprehensively and providing the theoretical basis and analysis tools for music signal processing. The PDF has been applied in music signal detection, identification, separation, classification and feature extraction and other areas. In this paper. the probability density functions of music signals are modeled as generalized Gaussian distribution (GGD). The research of its application in music source blind separation includes the following aspects:1. This paper does research the PDFs of music signals in depth. The histograms of different sources with different durations and the time-varied STFT component at each frequency band are made first, which is inspired by the implementation of GGD in speech probability modeling. By observing shapes of histograms, the probability density functions of music signals are modeled as GGD.2. Modeling the PDF of music signal with GGD. music blind separation is conducted in time domain. A way to determine the shape parameter is given, called dynamically parameter estimation method, which is used to solve the multi-channel music biind separation. In order to accelerate the convergence, down-sampling on the de-mixed signals is added to the dynamic parameter estimation ICA (DPEICA) method. Simulation results show that the DPEICA has good performance, and the algorithm is effective to music sources of both with similar distribution mixtures and hybrid mixtures. Meanwhile, the PDF of projection coefficient corresponding to basic functions of monaural blind separation in time domain is also modeled as GGD, which is proved effective by experiments.3. Generalized Gaussian distribution is applied in frequency domain music blind separation. The PDF of the time-varied STFT components at each frequency band is modeled as GGD, dealing with the time domain convolution mixture in frequency domain. In order to eliminate the uncertainties of permutation between each frequency band, this paper presents the permutation algorithm based on complex average correlation and simulation results show that it is effective and stable. Finally, permutation algorithm based on complex average correlation and GGD are combined to modify the traditional convolution blind separation in frequency domain, simulation results show that the algorithm has better performance.