Research On Underdetermined Blind Source Separation Method For Sparse Optimization Of Speech Signal

Blind source separation(BSS)is a key technology of blind signal processing,which is widely used in speech processing,image processing,medical signal decomposition,mechanical fault detection and other fields.With the rapid development of artificial intelligence,blind source separation(BSS)plays an important role in the front-end processing of intelligent speech system,such as speech recognition,speech enhancement and source localization.In practical engineering,the underdetermined blind source separation(UBSS)method for speech signals is studied in this paper.Based on the idea of sparse component analysis(SCA),in the classical two-step framework,this paper uses a variety of signal sparse transformation methods,optimization methods and compress sense(CS)theory to achieve underdetermined speech blind separation.Among them,the sparse optimization method of speech signal and the source signal recovery method are the research focus.In this paper,different blind separation algorithms for underdetermined speech are studied from the two sparse optimization directions of time-frequency transform and dictionary learning(1)This paper studies the sparse representation of non-stationary and stationary speech signals using different time-frequency transform methods,and proposes a greedy sparse optimization algorithm in frequency domain for source signal recovery of underdetermined speech blind separation.In this algorithm,greedy optimization is introduced into the source recovery method of sparse component analysis to realize underdetermined blind source separation of different kinds of speech signals.Compared with the shortest path method,the proposed algorithm can improve the separation performance of more than two mixed signals;compared with the smooth L0 norm algorithm,the proposed algorithm can effectively improve the separation performance of speech blind signals with near arrival direction.The simulation results show that the proposed algorithm has a wider application range while maintaining the separation quality.(2)Based on the existing dictionary domain sparse optimization methods,an improved source signal recovery algorithm based on KEMD-SimCO dictionary learning is proposed to enhance the sparse ability of sparse base dictionary and improve the performance of source signal recovery.The improved method uses empirical mode decomposition(EMD)to enhance the sparsity of the signal,and takes the K-means empirical mode decomposition(KEMD)dictionary as the initial Dictionary of the existing SIMCO dictionary learning method,which improves the dictionary sparsity ability.(3)Considering that different segment lengths have different effects on the recovery of speech signals,a novel improved source signal recovery algorithm based on dynamic SimCO dictionary learning is proposed.On the basis of sparse features of speech signals obtained by existing SimCO dictionary learning algorithm,the idea of steepest descent optimization is used to dynamically change the segment length of the signal to obtain the globally optimal recovery signal,so as to improve the efficiency of under attribute Performance of blind source separation.Simulation results show that the proposed KEMD-SimCO and dynamic SimCO methods can achieve source signal recovery successfully.Compared with the existing dictionary learning methods,the improved dynamic SimCO dictionary learning algorithm can more fully mine the sparse characteristics of speech in the dictionary domain,and improve the quality of speech recovery while ensuring high efficiency.