Researchers introduce a gating-based Dynamically Slimmable Network (DSN) aimed at further reducing the complexity of lightweight speech enhancement models. The design combines static and dynamic components, with gating used to make the network dynamically slimmable during operation.
The abstract describes the DSN as architecture-independent, though the released text cuts off before explaining how that property is achieved. The paper's title also mentions metric-guided training, but the abstract does not spell out which metric is used or how it guides the training process.
Because this is a single source, there are no other findings to compare. The preprint is available on arXiv under identifier 2510.11395.