A DOA-Based Semi-Blind Extraction For Underdetermined Convolutive Speech Mixtures

Speech signals recorded in natural environment are usually in the form of underdetermined convolutive mixtures, containing multipath speeches and noises, and the number of source signals is greater than that of receivers. Separation of target speech signal from such mixtures has been recognized as a major problem. Blind source separation (BSS) is a signal processing technique developed in the late1980s, and has shown its advantages in solving the separation problem of underdetermined convolutive speech mixtures.However, due to the inherent intractability of underdetermined and convolutive problem, the existing algorithms for underdetermined convolutive speech separation do not work well. Among them, some masking algorithms have been proposed under assumption of signal’s sparsity. Most of these methods are blind algorithms which use little prior information. In consideration of the location of the source signals is known in many practical situations, this thesis explores semi-blind BSS methods which can use spatial prior information in order to enhance the separation performance and extract the interesting speech signals.To this end, this paper mainly do the following three aspects:(1) Studied an existing time-frequency masking algorithm based on complex vector Hermitian angle, constructed a reference vector according to the specific speech’s direction of arrival (DOA) information, and then proposed a semi-blind extraction algorithm. The experimental results based on simulated and real speech mixtures showed that the semi-blind extraction algorithm can extract an interesting speech with improved performance.(2) Presented a semi-blind separation of speech signals by outputting all masks based on semi-blind extraction algorithm. For the permutation problems between adjacent frequency bins, an improved mask clustering reordering algorithm combined with correlation maximum was proposed. Separation experiment resultsusing simulated and reall speech mixtures showed that performance of the proposed semi-blind algorithm is higher than that of the original blind algorithm.(3) Since the number of source signals may be unknown in practice, this thesis integrated multiple cluster validity indices to improve the estimation performance. The experimental results showed that the method can determine a better clustering number for Hermitian angles in each frequency bin.