Scholarlink

<p>Transformer architectures have found widespread application in sound event detection (SED). However, due to its high computational complexity, most models employ a fixed context window, which hinders the modeling of long-range dependencies beyond the window boundaries. Mamba alleviates the limitation of fixed attention windows, but its inference remains constrained by sequential scanning, which limits parallelization efficiency. In this study, we propose NSM2-SED, a non-scanning structured state duality based Mamba-2 architecture for SED. The mean teacher (MT) model is used as the baseline system. Initially, we pre-train the NSM2 module on AudioSet and run it in parallel with a CNN-based feature extractor to capture a broader spectrum of audio features. Subsequently, a two-stage training strategy is adopted: in stage 1, the NSM2 module is frozen, while the CRNN is trained for effective component interaction; in stage 2, all parameters are fine-tuned jointly. Our experiments demonstrate that, NSM2-SED achieves performance comparable to, and in some cases surpasses, that of Transformer-based models, while requiring fewer parameters. And the proposed method achieved 0.510/0.762 PSDS1/PSDS2 in the DESED dataset, significantly outperforming the baseline results of 0.385/0.593. The source code has been released at <a href="https://github.com/lchenglong789-cell/NSM2-SED.git">https://github.com/lchenglong789-cell/NSM2-SED.git</a>.</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

NSM2-SED: a non-scanning Mamba-2 model for sound event detection

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月15日

查看全文
  • Chenglong Liu,
  • Shuguo Yang,
  • Wenwu Wang

摘要

Transformer architectures have found widespread application in sound event detection (SED). However, due to its high computational complexity, most models employ a fixed context window, which hinders the modeling of long-range dependencies beyond the window boundaries. Mamba alleviates the limitation of fixed attention windows, but its inference remains constrained by sequential scanning, which limits parallelization efficiency. In this study, we propose NSM2-SED, a non-scanning structured state duality based Mamba-2 architecture for SED. The mean teacher (MT) model is used as the baseline system. Initially, we pre-train the NSM2 module on AudioSet and run it in parallel with a CNN-based feature extractor to capture a broader spectrum of audio features. Subsequently, a two-stage training strategy is adopted: in stage 1, the NSM2 module is frozen, while the CRNN is trained for effective component interaction; in stage 2, all parameters are fine-tuned jointly. Our experiments demonstrate that, NSM2-SED achieves performance comparable to, and in some cases surpasses, that of Transformer-based models, while requiring fewer parameters. And the proposed method achieved 0.510/0.762 PSDS1/PSDS2 in the DESED dataset, significantly outperforming the baseline results of 0.385/0.593. The source code has been released at https://github.com/lchenglong789-cell/NSM2-SED.git.