To achieve more accurate sound event detection and localization in complex environments characterized by source overlap and strong noise interference, this study first conducts comparative experiments using multiple acoustic features as inputs to a deep learning network. The results indicate that LGAM (logarithmic Gamma spectrum) provides a richer representation of both high- and low-frequency information, and is therefore adopted as the input feature. Furthermore, considering the inherent temporal characteristics of acoustic signals, a novel attention-based architecture, termed ACANet (asymmetric convolutional attention network), was proposed. ACANet employs asymmetric convolution kernels to enhance the learning of temporal dependencies and integrates CCAM (coordinate channel attention module) to improve the modeling of channel and spatial information. Experiments conducted on the TAU-NIGENS Spatial Sound Events 2020 dataset demonstrate that, compared with the baseline, the proposed method achieves overall performance gains, reducing ER (error rate) by 5%, decreasing LE (localization error) by 1.9°, and improving F1 score and FR (frame recall) by 3.1% and 1.4%, respectively. These findings confirm the effectiveness and application potential of the proposed approach for enhancing SELD(sound event localization and detection) performance in complex acoustic environments.
SU Rui-xuan, ZHANG Lin-ke, YU Yong-sheng, GE Dong-yuan.
Sound Event Localization and Detection Based on Feature Selection and Novel Attention Network[J].
Science Technology and Engineering, 2026, 26(10): 4288-4298 DOI:10.12404/j.issn.1671-1815.2502557
声音事件定位与检测(sound event localization and detection,SELD)是指对复杂环境下的声音事件进行识别检测,并进一步估计声音事件相对于麦克风位置的空间方位,即把声音事件检测(sound event localization,SED)和波到达方向(direction of arrival,DOA)这两个独立任务联系起来,实现对声音事件时空场景的描述。SELD技术凭借其空间感知能力,已形成跨领域应用生态:在安防领域支撑智能声学监控系统的多目标追踪;赋能智能机器人的动态环境感知与避障决策;实现工业设备故障声源的精准定位与早期预警等。
基线SELDNet为DCASE2019比赛主办方提出的一个卷积循环神经网络(CRNN)模型,其结构如图4所示。SELDNet网络模型学习音频的Log-Mel时频域特征和GCC-PHAT时差信息特征实现声音事件的检测与定位。CNN部分由4个基本卷积神经网络层组成:①卷积Conv;②批归一化BN;③ReLU激活;④最大池化;如图5虚线框内所示。RNN部分则采用两个双向门控循环单元(bidirectional gated recurrent unit, BiGRU)组成。
ShiWei, WuJian. A multi-attribute evaluation method for low-altitude security system solutions based on DEMATEL and TOPSIS[J]. Scienfice Technology and Engineering, 2025, 25(14): 6109-6117.
MaGuohao, NiuChangliu, WangYang. Research on sound source localization algorithm based on neural Network[J]. Digital Technology and Application, 2021, 39(7): 106-109.
GaoChunyan, LaiGuangjin, LuXiaoling, et al. Review of metho-ds based on CNN for mobile robot sound source localization[J]. Science Technology and Engineering, 2024, 24(7): 2617-2624.
[9]
AdavanneS, PolitisA, VirtanenT. Dir-ection of arrival estimation for multiple sound sources using convolutional recurrent neural network[C]// 26th European Signal Processing Conference (EUSIPCO). Piscataway, NJ: IEEE, 2018: 1462-1466.
[10]
ButkoT, GonzálezF, SeguraC, et al. Two-source acoustic event detection and localization: online implementa-tion in a smartroom[C]// 19th European Signal Processing Conference. Piscataway: IEEE, 2011: 1317-1321.
[11]
ChakrabortyR, NadeuC, Sound-model-based acoustic-source localization using distributed microphone arrays[C]// IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). Piscataway: IEEE, 2014: 619-623.
[12]
HirvonenT. Classification of spatial audio location and content using convolutional neural networks[C]// Audio Engineering Society 138th. New York: Audio Engineering Society, 2015: 1857-1861.
[13]
HeW P, MotlicekP, OdobezJ M. Deep neural networks for multiple speaker detection and localization[C]// IEEE International Conference on Robotics and Automation (ICRA). Piscataway: IEEE, 2018: 74-79.
[14]
AdavanneS, PolitisA, VirtanenT. A multi-room re-verberant dataset for sound event localization and detection[J]. arXiv Preprint arXiv: 1905.0854, 2019.
[15]
ChenZ Y, HuangQ H. GLFE: global-local fusion enhancement for sound event localization and detection:DCASE2022[R]. Challenge: Technical Report, 2022.
[16]
Naranjo-AlcazarJ, Perez-CastanosS, FerrandisJ, et al. Sound event kocalization and detection using squeeze-excitation residual CNNs[J]. arXiv Preprint arXiv: 2006.14436, 2020.
[17]
CaoY, IqbalT, KongQ Q, et al. An improved event-independent network for polyphonic sound event localization and detection[C]// IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). Piscataway: IEEE, 2021: 885-889.
PuZijun, ZhangShouming. Research on sound event localization and detection algorithm based on feature fusion and transformer model[J]. Computer Engineering and Science, 2023, 45(6): 1097-1105.
XuChundong, LiuHao, MinYuan, et al. Sound event localization and detection based on dual attention[J]. Computer Engineering and Applications, 2023, 59(19): 99-105.
SongQingjun, JiaoShouyue, JiangHanyan, et al. Coal gangue audio classification method based on improved EfficientNet[J]. Journal of Mine Automation, 2025, 51(1): 138-144.
[24]
LiuK, YuS, LiuS. An improved InceptionV3 network for obscured ship classification in remote sensing images[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2020, 13: 4738-4747.
LiQingsheng, MiaoNan, ZhangXin, et al. Image recognition of maize disease based on asymmetric convolutional attention residual network and transfer learning[J]. Science Technology and Engineering, 2021, 21(15): 6249-6256.
[29]
HuJ, LiS, SunG, et al. Squeeze-and-excitation networks[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2018: 7132-7141.
[30]
HouQ, ZhouD, FengJ. Coordinate attention for efficient mobile network design[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2021: 13713-13722.
[31]
PolitisA, AdavanneS, VirtanenT. A dataset of reverbe- rant spatial sound scenes with moving sources for sound localization and detection[J]. arXiv Preprint arXiv: 2006.01919, 2020.
[32]
CaoY, KongQ, IqbalT, et al. Polyphonic sound event detection and localization using a two-stage strategy[J]. arXiv Preprint arXiv: 1905.00268, 2019.