Investigation Of Bayesian Inter-Correlation Block Sparsification And Its Learning Method Of Speech Sound Signals Based On Acoustic Features

With the deep integration of artificial intelligence technology and traditional manufacturing industry,easy and accessible speech interaction with intelligent devices has become an inevitable trend in the development of intelligent technology;however,natural intelligent speech interaction technology faces serious challenges and difficulties due to the strong time-varying,broadband,background noise sensitive,propagation pattern and spatial location coupling characteristics of speech.How to improve the speech interaction capability,i.e.the efficient sparsification method of speech acoustic signal and related technology,has become a key scientific problem that needs to be solved.Thus,based on an in-depth study of speech acoustic features and Bayesian sparsification,a Bayesian inter-correlation block sparsification and learning method for speech acoustic signals based on acoustic features is proposed to improve the efficient sparsification of speech acoustic signals.In addition,this research is supported by two National Natural Science Foundation of China(NSFC)projects(61871447,61671262).Main research contents.(1)The correlation and block sparse acoustic features of speech acoustic signals are investigated,and the corresponding characterization parameters are given.The sparse Bayesian learning method is also discussed in depth,and the sparse Bayesian learning model of acoustic signals and the regularized solving algorithm of the optimal sparse solution are discussed in detail.(2)Based on the above analysis,the signal sparsification mechanism based on the two loud features of inter-correlation and block sparsity of the speech signal is investigated,the promotion of correlation and block sparsity on efficient signal sparsification and the quantitative expression of their effects are clarified,and the signal block classification criterion,block sparsity,inter-block and intra-block signal correlation evaluation covariates of the speech acoustic signal are constructed.(3)Based on the above-mentioned covariates and interaction mechanisms,a Bayesian correlation block sparsification learning method for speech acoustic signals based on acoustic features is proposed.A feature penalty matrix is constructed to fuse the intercorrelation and block sparsity to promote solution sparsity in the feature space,and a Bayesian sparse regularization mathematical model is thus constructed to fuse the two loud features,and the optimal sparse solution is derived quickly by convex optimization.(4)The key parameters and their effects and the core algorithm are numerically analysed and experimentally explored by simulation analysis and experimental validation.In addition,the efficiency and reliability of the method are further verified in terms of solution sparsity,reconstruction accuracy and computational speed by comparing the method with traditional methods,providing theoretical method support for the application of efficient sparsification of speech signals.