The Research Of Speech Blind Separation Algorithm

Speech blind separation algorithms are derived from the blind source separation technology. Blind source separation technology is a new research filed in last decade of the 20th century. The goal of the blind source separation is to recover the original sources given only the observed mixtures.So far, algorithms of blind source separation can be divided into two branches: one is multi-channel blind source separation algorithm; Another is the single blind source separation algorithm.In this paper, the multi-channel blind source separation algorithm mainly uses Independent Component Analysis (ICA) method, for example, natural gradient algorithm, information maximization algorithm, maximum signal-to-noise ratio, global optimal property, flexible ICA algorithm and so on; the single-channel blind source separation algorithm mainly uses Computational Auditory Scene Analysis (CASA) method, for example, Wang-Brown model and Hu-Wang model.This paper mainly elaborates the basic principle of the ICA model with instantaneous and noiseless and linear mixed, and describes the basic principle of the adaptive natural gradient algorithm and batch natural gradient algorithm in the framework of the model, and implements the two algorithms. On the basic of the fixed step-size batch natural gradient algorithm, this paper proposes a novel variable step-size batch natural gradient algorithm. The simulated results show, comparing with fixed step-size algorithm, the variable step-size algorithm has much higher convergence speed and a smaller steady state error.This paper describes the basic principle of the information maximization algorithm, the basic information maximization algorithm is established by the sigmoid transform function, the proposed information maximization algorithm is based on the hyperbolic tangent, the simulated results indicate, comparing with sigmoid function, the hyperbolic tangent transform function algorithm has much higher convergence speed and a smaller steady state error.This paper research into blind separation algorithms based on maximum signal-to-noise ratio and global optimal property, the simulated results indicate that two algorithms have good separation performance.This paper describes the basic principle of flexible ICA algorithm, concept of the flexible ICA algorithm is derived from generalized gaussian distribution of the natural gradient framework. To estimate gaussian exponent, this paper proposes an improved flexible ICA algorithm. According to kurtosis of the generalized gamma distribution, this paper proposes a flexible ICA algorithm based on generalized gamma distribution, the simulated results indicate, comparing to the basic flexible ICA algorithm, the improved flexible ICA algorithm has much higher convergence speed and a smaller steady state error, and the steady state error of the flexible ICA algorithm based on generalized gamma distribution is little greater than the others algorithm, but this algorithm can also has good separation performance, and its convergence speed is higher than the others algorithms.This paper elaborates also the basic principle based on the Wang-Brown CASA model and Hu-Wang CASA model, and achieves application speech enhancement based on two models. In this paper, enhancement performance of two models is evaluated by segment Signal-to-Noise and similitude coefficient of the source signal and recovered signal. The simulated results indicate that enhancement performance of the Hu-Wang CASA model is better than Wang-Brown CASA model, Hu-Wang model also makes up shortage of that Wang-Brown model can not carry on the very good separation in the high frequency areas of speech.