Research Of Multiple Sound Source Localization And Separation Based On Signal Sparsity

Spatial audio where spatial feature extraction and sound field reproduction are very important,helps to create immersive virtual soundscapes.For current techniques,each source objects in the sound scene can be recorded through an exclusive microphone.However,it does not work in real-time applications,where there are only recording mixtures can be provided for processing.Considering this situation,multiple sound source localization and separation become very crucial to extract the original signals and corresponding spatial information.This work focus on the research of multiple sound source localization and separation based on the sparsity of signals,aiming at developing efficient localization and separation approaches to adapt various acoustic conditions.To overcome the drawbacks of existing approaches for source localization and separation,we focus on developing generalized approaches for multiple source localization and separation by using less number of microphones but with better performance.The researches of the dissertation include the following aspects:First of all,a multiple sound source localization and counting method based on a relaxed sparsity of speech signal is presented.After establishing an effective measure,the relaxed sparsity of speech signals is investigated.According to this relaxed sparsity,we can obtain an extensive assumption that “single-source” zones always exist among the soundfield microphone signals,which is validated by statistical analysis.Based on“single-source” zone detecting,the proposed method jointly estimates the number of active sources and their corresponding DOAs by applying a peak searching approach to the normalized histogram of estimated DOA.The evaluations reveal that the proposed method achieves a higher accuracy of DOA estimation and source counting compared with the existing techniques.Next,a single source bin(SSB)based multiple source localization scheme is proposed.First,a “DOA convergence” assumption is proposed,which means that if most of the time-frequency(TF)bins in a TF zone are derived from only one source — defined as SSBs,the corresponding DOA estimates are relatively concentrated with a heavy density.This assumption is validated through statistical analysis by applying a quantitative measure of convergence.Accordingly,by applying the “DOA convergence” assumption,the detection of SSBs is converted to a clustering problem,K-means clustering and density-based spatial clustering of applications with noise(DBSCAN)algorithms are utilized to detect the SSBs.Experimental results demonstrate the localization accuracy of the proposed method outperforms the state-of-the-art localization approaches which are based on single source zone detection.Lastly,a BSS method for recovering multiple speech sources from sound fields recorded by a soundfield microphone.Such sparse components correspond to bins where only one speech source is active and are identified based on the inter-correlation among the mixture signals.Proposed is a “local-zone stationary” assumption,where the amplitude of a speech signal remains approximately constant within a small band of TF components.This assumption is validated through statistical analysis of a quantitative measure of stationary.Under this assumption,the non-sparse components are recovered by regarding sparse components as a guide.The final separated sources are obtained by combining the separated sparse and non-sparse components.Both objective and subjective evaluations show that the proposed method achieves better separation quality compared to some existing BSS approaches.