基于多随机森林的低信噪比声音事件检测  

Sound Event Detection at Low SNR Based on Multi-random Forests

作　　者：李应[1,2] 印佳丽 LI Ying;YIN Jia-li(College of Mathematics and Computer Science,Fuzhou University,Fuzhou,Fujian 350116,China;Key Lab of Information Security of Network Systems(Fuzhou University),Fuzhou,Fujian 350116,China)

机构地区：[1]福州大学数学与计算机科学学院,福建福州350116 [2]网络系统信息安全福建省高校重点实验室,福建福州350116

出　　处：《电子学报》2018年第11期2705-2713,共9页Acta Electronica Sinica

基　　金：国家自然科学基金(No.61075022);福建省自然科学基金(No.2018J01793)

摘　　要：论文针对各种背景声音中低信噪比声音事件的检测问题,提出把背景声音与声音事件混合,形成带噪声样本来训练分类器.在预处理阶段,使用基于经验模态分解与2-6级固有模态函数的投票方法,对背景声音与声音事件端点进行预测并估算信噪比.接着使用子带能量分布方法,提取声音数据的特征.最后,论文将背景声音与声音事件样本库中所有声音样本按照估算的信噪比相混合,生成混合声音特征训练多随机森林,用于低信噪比声音事件的检测.实验证实,所提出的方法可以用于各种声场景下低信噪比声音事件的检测,并能在信噪比为-5dB的情况下保持67. 1%的平均检测率.For sound event detection under various background noises at low SNR,this paper proposes a method that mixes the background noises with sound events into noisy samples to train classifiers.In the pre-processing stage,we use a voting method based on 2th to 6th intrinsic mode functions(IMFs)that generated from empirical mode decomposition(EMD),to detect the endpoint of sound events and estimate the SNR.Then subband power distribution(SPD)is used to extract features from audio data.Finally,we mix the background noise and all the sound event samples in the sound event database according to the estimated SNR,and then extract the noisy samples features to train multi-random forests(M-RF)for the detection of the sound events in low SNR environment.The experiment proves that the proposed method has the ability to recognize sound events in various acoustic scenes at low SNR,and can remain an average accuracy rate of 67.1%at-5dB

关 键 词：声音事件检测 信噪比 经验模态分解 子带能量分布 随机森林 

分 类 号：TP391.42[自动化与计算机技术—计算机应用技术]