Research On Speech Separation Based On Spatial Filtering

In real life, human beings can focus on some interested speech in a complex environment. The computer also can have this ability which is to extract one or more interested speakers' speech from numbers of speakers'speech by speech separation. Based on time domain and frequency domain, spatial domain is added to provide a new way of thinking for speech separation algorithm.The new way of thinking is that speakers'speech are separated according to the different positions of the speech. In this paper, we research on speech separation following the new idea.The main content of this dissertation can be summarized as followings:First, characteristics of speech signal and acoustic wave propagation are analyzed, and near and far field signal models of the microphone uniform linear array are researched based on traditional array signal processing.Second, high-resolution spectral estimation (such as MUSIC) and steered beamformer (such as MVDR) have poor noise proof. Due to the sparsity of array signal, sparse decomposition method is applied to sound source location to solve this problem. Coherent sources also can be located by this method. This sound source location method was proved to be effective via computer simulation. MP method is to find the best atom in atom library which is established via different orientation parameters by global search. So the algorithm is extraordinary complexity. To solve this problem MP method based on near-field sector division is used. The computer simulation proved that this algorithm greatly reduces the computational.So MP method is possible applied to real-time system.Third, near-field adaptive beamforming algorithm is researched on. According to the microphone array near-field model traditional capon beamforming algorithm is applied to near-field environment. However this algorithm is lack of robustness. So we research on near-field robust capon beamforming (RCB) algorithm to solve this problem. Near-field robust capon beamforming algorithm has robustness against both steering vector mismactches and finite-sample effects.Last, Combination of sound source localization algorithm and beamforming algorithm, we simulated the speech separation using matlab.