基于ERB响度特征的深度学习鸟鸣声识别  

Bird Song Recognition Based on ERB Loudness Feature and Deep Learning

作　　者：尹晨畅 许枫[1] 张纯[1] YIN Chenchang;XU Feng;ZHANG Chun(Ocean Aconstic Technology Center,Institution of Acoustics,Chinese Academy of Sciences,Beijing,100190,China;University of Chinese Academy of Sciences,Beijing,100049,China)

机构地区：[1]中国科学院声学研究所海洋声学技术实验室,北京100190 [2]中国科学院大学,北京100049

出　　处：《网络新媒体技术》2022年第2期25-32,共8页Network New Media Technology

基　　金：国家重点研发计划项目资助(编号:2017YFC1403501)。

摘　　要：为了适应现代化生态系统监测需求,提出了一种基于ERB尺度的响度特征的鸟鸣声识别方法,该方法先对信号进行分帧、加窗和傅里叶变换得到其短时功率,然后输入到外、中耳传递函数,最后将听觉域信号映射到临界带域获得时间-临界带的响度特征。根据八类鸟鸣声分别制作基于响度特征、时频特征和梅尔倒谱特征的数据集,输入到深度神经网络中进行了识别。实验结果表明:该方法与基于时频特征和基于梅尔倒谱特征的识别方法相比在保持了最高的平均识别率的同时极大地减少了识别时间和数据处理量。In order to adapt to the requirements of modern ecosystem monitoring,a method of bird song recognition based on ERB(Equivalent rectangular bandwidth)scale loudness feature was proposed in this essay.Firstly,the short-time power feature of the signal was obtained by frame segmentation windowing and Fourier transform,and then the short-time power feature was input to the external and middle ear transfer function.Finally,the auditory domain signal was mapped to the critical band domain to obtain the time-critical band loudness feature.Based on the characteristics of loudness feature,time-frequency feature and mel-frequency cepstrum feature,the data sets of eight kinds of bird songs were made and input into deep neural networks for recognition.The experimental results show that compared with the recognition method based on time-frequency feature and mel-frequency cepstrum feature,this method keeps the highest average recognition rate and greatly reduces the recognition time and data processing.

关 键 词：ERB尺度 短时功率 响度特征 深度神经网络 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] TN912.34[自动化与计算机技术—控制科学与工程] X835[电子电信—通信与信息系统]