The Study On Blind Separation Of Noisy Speech Mixtures

Humans can distinguish and follow the interested speech signal in several speakers’situation, which is a unique human ability for understanding. In the field of speech signal processing, how to separate each source speech signal from the observed speech mixtures, is an important problem. Blind source separation (BSS) is one of the main methods of Speech separation, which could separate original signals from some observed mixed signals without knowing any priori information on the all source signals. Generally, blind source separation methods are mostly carried out in the noise-free situation, but in the real environment, the speech signals can inevitably be affected by various noises, therefore it is of great value and significance to research speech separation in noisy environment.In this thesis, a novel blind separation algorithm of noisy speech is proposed based on improved spectral subtraction. First of all, the improved spectral subtraction algorithm is used to improved signal-to-noise ratio (SNR) of the noisy speech, and then to separate original signals from the noisy speech mixtures. The main work is organized as follow:First, the develop status of speech enhancement and its applications in speech separation is introduced. In the thesis, an improved spectral subtraction algorithm based on modified noise estimate and magnitude compensation is proposed, it can not only effectively remove the noise but also largely avoid source signals damaged, and this is of great help to signal separation after de-noising, can be largely avoid affecting the separation due to the destruction of the source signal.Second, the independent component algorithm (ICA) is used to get the de-noised mixtures separated. However, the FastICA algorithm based on negative entropy is sensitive to the initial value and has local maximum problem, and to solve the problem, a novel refined FastICA algorithm is proposed, this method is improved by combing Newton descent method and Modified fast independent component analysis (M-FastICA) to change the iteration mode, this will ensure separation effect while reducing the number of iteration, and also reduce the sensitivity of the initial value. To further enhance the accuracy of the speech separation algorithm, this method combines the distribution characteristics of the speech signal to choose different nonlinear function; in the end, it carry out the post de-noising of the separated signals, so as to further enhance the quality of separation of speech signal.Simulation results show that the proposed algorithm could achieve speed signal separating effect. The proposed method has a better separation performance, indicted by the similarity coefficient matrix and the minimum mean square(MSE); The algorithum’s complexity is also reduced that the number of iteration is decreased60%.