Scholarlink

<p>Multi-channel speech enhancement is essential for real-time applications requiring robust noise suppression and low latency. However, most existing U-Net-based approaches rely on parameter-heavy modules to achieve high performance, leading to excessive computational and memory costs that hinder deployment on resource-constrained devices. To address these limitations, we propose DB-CANet, a lightweight dual-branch dense network with time–frequency coordinate attention. The dual-branch architecture explicitly decouples spatial and spectral modeling, while depthwise separable convolutions (DSConv) and dense connection blocks (DCBs) mitigate parameter redundancy significantly. To further minimize complexity, two attention mechanisms–coordinate attention for time-frequency localization (CATFL) and efficient channel attention (ECA)–are integrated into DCBs, which allow the network to adaptively highlight critical time–frequency dependencies. Experiments on benchmark datasets demonstrate that the proposed DB-CANet achieves competitive enhancement quality with only 0.15M parameters, offering an effective balance speech quality and efficiency.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

DB-CANet: A Lightweight Dual-Branch Dense Network with Time–Frequency Coordinate Attention for Multi-Channel Speech Enhancement

  • 期刊论文

  • 发表日期：

    2026年3月6日

查看全文
  • Xinyue Zhou,
  • Yuankai Zhang,
  • Xiaoran Yang,
  • Gongping Huang

摘要

Multi-channel speech enhancement is essential for real-time applications requiring robust noise suppression and low latency. However, most existing U-Net-based approaches rely on parameter-heavy modules to achieve high performance, leading to excessive computational and memory costs that hinder deployment on resource-constrained devices. To address these limitations, we propose DB-CANet, a lightweight dual-branch dense network with time–frequency coordinate attention. The dual-branch architecture explicitly decouples spatial and spectral modeling, while depthwise separable convolutions (DSConv) and dense connection blocks (DCBs) mitigate parameter redundancy significantly. To further minimize complexity, two attention mechanisms–coordinate attention for time-frequency localization (CATFL) and efficient channel attention (ECA)–are integrated into DCBs, which allow the network to adaptively highlight critical time–frequency dependencies. Experiments on benchmark datasets demonstrate that the proposed DB-CANet achieves competitive enhancement quality with only 0.15M parameters, offering an effective balance speech quality and efficiency.