Sandglasset: A Light Multi-Granularity Self-attentive Network For Time-Domain Speech Separation PAPER LINE:https://arxiv.org/pdf/2103.00819v2.pdf less run time than GALR but i recommend SUDORMRF
thanks YoungloLee two issues： (1)do not share the same block in multi-granularity (2)control the size of parameters,we should use depth-wise convolution