Study On Speech Separation Algorithm Based On Subband Decomposition

IntroductionEmering multimedia applications such as hands-free communication and teleconferencing require sophisticated signal processing techniques to enhance the quality of speech signals. The degration of speech may happen in many possible ways. For example, speech may get corrupted because of the background noise generated by engine or wind in a car or fan in a room. In many instances, it is also desired to separate or "pick up" a speech signal recorded in the presence of other speakers. An example is the cocktail-party problems. Humans have a tremendous capability of focusing their attention on a speech signal in the presence of other speakers talking in the background. On the other hand, the front end of a speech-acquisition system lacks this capability, and proper signal processing should be applied to "extract" the desired speech signal from its mixture with interfering speakers. This problem can be described to the ability which can focus someone's talk in noisy environment.Some difficulties in speech blind separation are studied in this paper as followed: (1) Speech separation algorithms rely on the choice of nonlinear functions which is the estimination of probability distributions of speech sources. The appropriate nonlinear functions can enhance speed of convergence of separation algorithm. (2) To apply speech separation algorithm, generally speech source signals estimated must satisfy strong assumption that sources signals are mutually independ. When source signals are dependent, speech separation becomes diffcult.Laplace normal mixture distribution function is used as a statistical model for speech signal in this paper. It is known that the proposed speech separation algorithm attains fast convergence properties. When speech source signals are dependent, mixed signals can be decomposed to subband signals by subband decomposition. The demixing matrices in appropriate subband are used as demixing matrices of original mixing speech signal.1. Speech separation based on Laplace Normal Mixture Distribution Probability density function estimationTypical speech separation learning algorithms rely on the choice of nonlinear functions, the optimal form of which, depends on probability distributions of speech sources. Since probability distributions of sources are not known in advance in the algorithm, we count on the hypothesized density models for speech sources. The signal nonlinear function is favorable for the separation of speech signals since natural speeches are often modeled as Laplacian distribution. Therefore, Laplace normal mixture distribution function is used as a statistical model for speech signal in this paper. The nonlinear activation function deduced by Laplace normal mixture distribution is used in natural gradient separation algorithm. The proposed algorithm gives faster convergence and better performance.2. Subband DecompositionWe have mostly concentrated on the design and realization of single-input single-output digital filters. There are applications where it is desirable to separate a signal into a set of subbsnd signals occupying the whole Nyquist range. Each subband signal is processed according to character of subband signal. To this end, subbsnd filter banks play an important role in signal processing.Prototype Filter DesignThe low-pass prototype filter is a pivotally portion of subband filter banks. The design of the whole filer bank thus reduces to that of the prototype filter. A new objective function for optimization, a convex function of the parameter is proposed. The prototype filter can be designed by iteratively adjusting passband edge to minimize the objective function.Cosine Modulated Filter BanksThe filter banks are modulated versions of a low-pass prototype filter. The function to modulate prototype filter includes exponential function and cosine or sine function. The gained filter banks are DFT modulated filter banks and cosine modulated filter banks (CMFB) respectively. The CMFB can be simply designed and is easy to-be implemented. The outstanding advantage of CMFB is that during the design phase, where the optimization of the filter coefficients is conducted, the number of parameters to be optimized is very small since only the prototype filter has to be optimized.