摘要
针对现有的环境声事件识别方法对噪声的鲁棒性差以及街道环境声数据少的问题,提出一种改进的对数梅尔谱特征结合卷积神经网络的环境声事件识别方法。该方法将对数梅尔谱及其一阶差分系数和二阶差分系数构建为三维特征,使用卷积神经网络进行分类,提高了街道环境声事件识别的抗噪性能。采用自主设计的环境声采集设备收集了大量的街道环境声数据。实验结果表明,在不同环境声数据集下,该方法比常规的识别方法性能更优。此外,基于该方法提出了一套街道环境声检测系统,经实景测试,该系统的查全率、查准率和置信度分别为94%、87%和90.5%,相比于基于常规识别方法的检测系统具有更好的表现,进而验证了该检测系统的可行性。
Aiming at poor noise robustness of existing environmental sound event recognition methods and the lack of street environmental sound data.This paper proposes a method for recognizing environmental sound events based on improved log-Mel features combined with convolutional neural networks,firstly,it constructs the log-Mel and its first-and second-order into three-dimensional features,and then uses a convolutional neural network to classify,which effectively improves its anti-noise performance.Experimental results show that the proposed recognition method performs better than the conventional recognition methods under different environmental sound data sets.In addition,based on this method,a set of street environmental sound detection systems is proposed.After real-world testing,the Recall,Precision,and F1-measure of the proposed detection system are 94%,87%,and 90.5%,respectively.Compared with detection systems based on conventional methods,it has better performance,and which verify the feasibility of the proposed system.
作者
张留军
王玫
罗丽燕
ZHANG Liujun;WANG Mei;LUO Liyan(Provincial Ministry of Education Key Laboratory of Cognitive Radio and Signal Processing,Guilin University of Electronic Technology,Guilin 541004,China;College of Information Science and Engineering,Guilin University of Technology,Guilin 541006,China)
出处
《桂林电子科技大学学报》
2020年第5期411-417,共7页
Journal of Guilin University of Electronic Technology
基金
国家自然科学基金(61771151)
广西重点研发计划(2017AB08072)
广西自然科学基金(2016GXNSFBA38014)
中国博士后科学基金(2016M602921XB)
广西研究生教育创新计划(YCSW2019139)
桂林电子科技大学研究生教育创新计划(2019YCXS038)。
关键词
环境声事件识别
卷积神经网络
检测系统
environmental sound event recognition
convolutional neural network
detection system
