Separation Structure Speech Separation Model Fusing Conformer and NBC

Abstract The quality of speech separation affects the entire technology ecosystem. Aiming at problems low utilization local feature information, insufficient convergence speed, too many calculation parameters and long time in blind source view Transformer dual-path cyclic neural network, a model based on fusion Conformer NBC (Narrow-band NBC) is proposed. First, for information block replaced by recurrent network. It can improve speed features. Secondly, used to replace inter-block loop. simplifies vector similarity aggregation, reduces lots cost, computational complexity model. In experiments WSJ0-2mix [7] dataset WHAM [13] dataset, contrast with other models, faster effect better.