MSAG-Net: A Multi-Scale and Attention-Guided GRU Network for Snore Detection

To address the challenge of distinguishing between snoring and non-snoring sounds, this study proposes MSAG-Net, a deep learning model that integrates multi-scale convolution and attention mechanisms to improve snore detection accuracy and support early screening of sleep-related breathing disorders. The model employs a dual-scale convolutional structure to extract both local and global features from Mel spectrograms and incorporates the Convolutional Block Attention Module (CBAM) to enhance the representation of discriminative features. A channel-slicing strategy is used to convert the 3D spectrogram into a sequential format, which is then modeled by a two-layer GRU network for temporal dependency learning. Experiments using Mel spectrograms generated via Short-Time Fourier Transform (STFT) demonstrate that MSAG-Net outperforms baseline models without attention mechanisms in terms of accuracy and generalization. Ablation studies further confirm the effectiveness of each module, offering a robust and contactless solution for snore recognition.