Research On Sparse-Decomposition-Based Single-Channel Speech Separation Algorithm

Single-channel speech separation is a vital issue of speech separation, and sparse decomposition has provided a new way to solve such problems. Therefore, we put our research focuses on the sparse-decomposition-based single-channel speech separation algorithms in this dissertation.Using the theoretical results of sparse decomposition and compressed sensing (CS), which is developed on the basis of sparse decomposition, we propose several source-adaptive basis (or dictionary) construction algorithms first, and then on this basis, we propose several effective sparse-decomposition-based single-channel speech separation algorithms and CS-based single-channel speech separation algorithms, to enhance the target speech and suppress the interference speech. The main work and main contributions are described as follows:â‘ We derive the ideal quasi-KLT (QKLT) basis by diagonalizing the autocorrelation matrix of the speech source, and prove that, all the speech sources are sparse in their ideal QKLT bases. However, in some real-world applications, the ideal QKLT bases cannot be obtained exactly, therefore, we propose to construct two types of template-matching QKLT bases, which are called nonhomogeneous linear mean square estimation (NLMSE) template-matching QKLT basis and orthogonal matching pursuit (OMP) template-matching QKLT basis. Simulation results demonstrate that the characteristics of voiced speech signals in their template-matching QKLT bases are similar with those in their ideal QKLT bases.â‘¡On the basis of the work described inâ‘ , we propose the single-channel speech separation algorithms based on QKLT bases.a. We prove that, based on the ideal QKLT bases, all the sources can be perfectly separated from a single mixture by l 0 optimization.b. Since the sources are unknown before speration, we cannot obtain the ideal QKLT bases actually. Therefore, we propose to perform single-channel speech separation based on the NLMSE template-matching QKLT basis and OMP template-matching QKLT basis separately, which are constructed from the mixture by improving the two template-matching QKLT basis construction algorithms described above. It is observed that our proposed methods perform better than the method exploiting independent component analysis basis functions and the improved computational auditory scene analysis based method using shape analysis.â‘¢We study on CS-based single-channel speech separation algorithms.a. Motivated by the similarity between the mathematical model adopted in speech separation and CS, we attempt to recover the speech sources using a CS appoarch, and propose a single-channel speech separation algorithm based on CS and K-SVD. Simulation results demonstrate that our proposed algorithm has stable performance.b. Based on the sparsity of speech in the DCT domain, we propose an inter-frame and intra-frame adaptive speech CS method using speech energy, to further reduce the speech sampling rate based on CS theory, for the further research on single-channel speech separation based on CS measurement characteristics. Simulation results demonstrate that our algorithm performs better than the bayesian compressed sensing.