Voice signal processing system and method based on layered inattention model

The invention provides a voice signal processing system and method based on a hierarchical inattention model, and the system comprises a voice preprocessing embedding module which is used for obtaining the voice information of a user, and extracting the feature vector of the voice information of the user; the hierarchical inattention module comprises a plurality of inattention hierarchies, each hierarchical structure comprises a plurality of AFFormer units, each AFFormer unit comprises a token mixer and a channel mixer, and the token mixer comprises a plurality of parallel depth separable convolution branches and is used for processing the received feature vectors to obtain token information output; and the channel mixer comprises a nonlinear gating branch and a linear branch which are parallel to each other and are used for respectively processing the received token information, obtaining an optimal gating signal characteristic value according to processing results of the two branches, and obtaining target i