Under-determined Blind Speech Source Separation Based On The Space Geometric Constraints In The Convolutive Case

Blind source separation(BSS) is the process of estimating the source signals from the observed signals, which is their mixtures, without any prior knowledge about the sources or the mixing process. With the deepening of research, it has been applied in a wide field, such as biomedical signal processing, image recognition,wireless communications, radar and sonar signal processing, speech processing and so on. In practical application, the source signals may be estimated from less observed signals, i.e., the number of the observed signals is less than that of the sources, which is named as Under-determined blind sources separation(UBSS). For the UBSS problem, the traditional blind source separation algorithms, which based on independent component analysis, cannot be directly used for the source separation due to the mixing system is irreversible. Generally, the sparse component analysis could be used to resolve the UBSS problem. However, the mixing process of speech is the convolution mixing process in the real world, so the sparsity of the signal is restricted. According to the method of perception for the speech signal by human,which used the learning system and the visual information, an algorithm is proposed in this dissertation to solve the problem of under-determined convolutive blind source separation. This algorithm is based on state-space model by using the constraints of sources’ spatial location. The main contributions of this dissertation are summarized below:For the convolution mixing model, the temporal correlation within the speech signal can be used to improve the performance of separation. Then motivated by the fact that the mixing filter effected by the spatial location of the sources and microphones, an algorithm based on state-space model is proposed. The estimations of sources are obtained by using Kalman filtering on the state-space model. In the convolution case, the permutation problem is an another problem which needs to be solved. In this dissertation, the initial value of the model parameters is obtained by using the spatial location of sources and microphones, which is obtained by the visual information. According to the spatial location, the mixing filters could be initializedin the same order across all frequency bins. And the order would not be changed during the process, which means that the order of each column vector of the mixing matrix is as same as that of the initial value. Because of that, the permutation problem could be avoided. Experimental results over reverberant synthetic mixtures and live recordings of speech data show the effectiveness of the proposed algorithm.