Improved Wave-U-Net Network for Speech Enhancement in Ocean Noise Environment

In recent years, speech enhancement methods based on deep neural networks have gradually emerged, but they still face challenges such as noise diversity and unique signal characteristics in the Marine environment. To solve these problems, this paper proposes an improved Wave-U-Net network, which combines multi-scale feature extraction and channel attention mechanism to enhance feature interaction and expression. In the training process, the loss function of integrated time domain and frequency domain errors is used, and the multiscale feature extraction module and dual attention mechanism are used to dynamically assign weights to focus on key features and suppress noise. The experimental results show that the proposed network PESQ reaches 3.37 and STOI reaches <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 3. 5 2 \%}$</tex>. This model can effectively remove ocean noise and retain key features of speech, significantly improve the performance of speech enhancement, and provide a new solution for voice communication in complex Marine environments.