Multi-scale Local and Global Context-aware Network for Speech Enhancement in Insect Noise Environments

In the summer and autumn seasons, insect chirping (such as cicadas) emerges as the predominant environmental noise interference in outdoor audio data collection. Although existing speech enhancement models have made progress, they often fail to fully utilize multi-scale information and ignore feature interactions between scales. To address these limitations, this paper introduces a Multi-scale Local and Global Context-aware Network (MLGNet) for solving insect noise interference. MLGNet is based on UNet with Local-Global Scale-Aware (LGSA) module, aiming to learn a wider range of multi-scale contextual features. LGSA consists of three key components. Specifically, Multi-scale Global-based Attention (MGA) module aggregates multi-scale information from different channel levels for global feature modeling and encodes causal position information. The Multi-scale Interactive Convolution (MIC) module achieves crossscale interaction modeling of local features through dual scale cross learning. Meanwhile, Context-aware Enhancement Fusion (CEF) module adaptivly learns of local and global feature weights and enhances nonlinearity, promoting high-quality dynamic fusion of features. The quantitative evaluation of the Insect Noise dataset and the Voice Bank+DEMAND dataset demonstrate that MLGNet outperformed other models and achieved excellent results in objective evaluation metrics.