Research Of High-efficiency Blind Source Separation Algorithm

Many algorithms have been dedicated to working on how to separatehigh-quality signals as the applications of blind source separation extending.Problems to be solved by this paper is separating the high-quality signals whileimproving the speed and reducing the computational complexity of the algorithm.Achieving efficient blind separation algorithm is mainly from the following twoaspects.For the instantaneous mixture model of the blind source separation, this paperpresents the method of using the amplitude matrix obtained in frequency domain toobtain the separation matrix and then separate the mixture signals. At this time, thefrequency the amplitude of the observation signals must be sufficiently accurate.Therefore, this paper introduces a method of spectrum correction as this method hashigh correction precision for the discrete non-intensive spectral. So, taking advantageof this method to correct the spectrum for the mixture signals, and then obtain thefrequency and amplitude accurately. To obtain the separation matrix, we should havea clustering of the amplitude matrix. That is: bringing the amplitude vectors from thesame source together. Due to the uncertainty of the mixing matrix, the clusteringalgorithm can‘t adopt the Euclidean distance but the angle between two vectorsclustering. Based on the above algorithm, we can obtain the separate matrix and thesignals. Experiments show that the proposed method of basing on the amplitudematrix in frequency domain has high accuracy and low complexity, and also save thecomputation time.For the convolution mixed model of blind source separation, this paper proposedan algorithm based on extracting the sparse time-frequency characteristic patterns.First, the mixture speech signals are transformed to time-frequency domain, as thesparse characteristic simplifies the blind separation algorithm and is convenience toremove the noises. In the meanwhile, highly redundant of STFT framework and themasking of ears are theory basis of removing the redundant information and reducingthe number of clusters of the characteristic patterns and also computing speed of thealgorithm. The removal redundant information is not the primary component formedthe source signals. So, it will not affect the quality of the separation signals. This paper can reduce94-95percent of the characteristic patterns, so it can save a lot ofcomputations and accelerate the calculation speed. Based on the a priori transmissionmodel, this paper obtains the characteristic patterns by standardized, which makes theclustering has nothing to do with the frequency. So, we can solve the orderuncertainty of the convolution mixed model. Finally, we can separate the signals withT-F masking.Through the study of the two models of blind source separation, we can see thatthe quality of the separated signals is guaranteed and then the complexity of thealgorithm is still worth studying. This can save a lot of resources and memory, andaccelerate the speed of the signal separation. It can also save times, and have practicalengineering significance. Experiments show that the algorithm proposed in this papercan achieve high quality of separated signals and low complexity of the algorithm.