Scholarlink

<p>Sound event detection (SED) aims to classify sound events in complex acoustic environments and determine their temporal occurrences within an audio clip. Many supervised SED methods are trained based on strongly labeled data, but obtaining such annotations in complex environments is often challenging. This paper proposes an environmental sound event detection method based on weakly labeled learning. The proposed model employs dynamic multi-scale convolution to enlarge the receptive field of the model and adaptively assign weights to convolutional kernels at different scales based on the input, thereby enhancing feature representation. Global dynamic attention is introduced to identify the important features and capture the global dependencies of key time-frequency information. Global weighted ranking pooling is used to model the classification mapping. Our method achieved an average improvement of 6% over the baseline in ablation experiments. We utilized the UrbanSound8K dataset as the sound event to construct non-overlapping datasets and overlapping datasets under different signal-to-noise ratios by using the multi-instance learning method. In the multi-label audio classification task, our method outperforms the state-of-the-art classification model by 2% in both F1-score and AUC, while reducing the number of parameters by more than 90%. For the SED task, the proposed method achieves a 2% improvement in F1-score, a 3% increase in AUC, and a 6% reduction in ER compared to the best-performing weakly labeled learning SED model, along with a 50% reduction in model parameters. These results demonstrate that our method can achieve state-of-the-art performance with significantly smaller model size. In the generalization experiments conducted on the DCASE 2024 Task 4 DESED dataset, our method achieves state-of-the-art performance, with an overall improvement of 4%, confirming its effectiveness and robustness. These results highlight the potential of our approach for rapid and reliable sound event detection, particularly in resource-limited settings. Our approach offers an effective solution for environmental noise monitoring, which could facilitate broader deployment in real-world applications.</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

Weakly Labeled Environmental Sound Event Detection Based on Dynamic Multi-scale Convolution Attention

  • 期刊论文

  • 发表日期：

    2025年10月14日

查看全文
  • Baojun Chen,
  • Jianxin Peng

摘要

Sound event detection (SED) aims to classify sound events in complex acoustic environments and determine their temporal occurrences within an audio clip. Many supervised SED methods are trained based on strongly labeled data, but obtaining such annotations in complex environments is often challenging. This paper proposes an environmental sound event detection method based on weakly labeled learning. The proposed model employs dynamic multi-scale convolution to enlarge the receptive field of the model and adaptively assign weights to convolutional kernels at different scales based on the input, thereby enhancing feature representation. Global dynamic attention is introduced to identify the important features and capture the global dependencies of key time-frequency information. Global weighted ranking pooling is used to model the classification mapping. Our method achieved an average improvement of 6% over the baseline in ablation experiments. We utilized the UrbanSound8K dataset as the sound event to construct non-overlapping datasets and overlapping datasets under different signal-to-noise ratios by using the multi-instance learning method. In the multi-label audio classification task, our method outperforms the state-of-the-art classification model by 2% in both F1-score and AUC, while reducing the number of parameters by more than 90%. For the SED task, the proposed method achieves a 2% improvement in F1-score, a 3% increase in AUC, and a 6% reduction in ER compared to the best-performing weakly labeled learning SED model, along with a 50% reduction in model parameters. These results demonstrate that our method can achieve state-of-the-art performance with significantly smaller model size. In the generalization experiments conducted on the DCASE 2024 Task 4 DESED dataset, our method achieves state-of-the-art performance, with an overall improvement of 4%, confirming its effectiveness and robustness. These results highlight the potential of our approach for rapid and reliable sound event detection, particularly in resource-limited settings. Our approach offers an effective solution for environmental noise monitoring, which could facilitate broader deployment in real-world applications.