UL-UNAS: Ultra-Lightweight U-Net for Real-Time Speech Enhancement via Neural Architecture Search

Researchers have introduced UL-UNAS, an ultra-lightweight U-Net optimized through Neural Architecture Search (NAS) for real-time speech enhancement. The model is designed for low-footprint devices, addressing the growing demand for compact models in this field. The team explored various efficient convolutional blocks within the U-Net framework, identifying promising candidates. They then enhanced these blocks with two boosting components: a novel activation function called affine PReLU and a causal time-frequency attention module. By leveraging NAS, they discovered an optimal architecture within a carefully designed search space. UL-UNAS reportedly outperforms the latest ultra-lightweight models, as stated in the paper. The work is detailed in a paper available on arXiv with identifier 2503.00340, announced as a cross-type submission.