基于滤波器组和残差网络的环境声音识别算法  

Environmental sound recognition algorithm based on filter bank and residual network

作　　者：刘亚荣[1,2] 于顼顼 谢晓兰[1,2] LIU Ya-rong;YU Xu-xu;XIE Xiao-lan(Guangxi Key Laboratory of Embedded Technology and Intelligent System,Guilin University of Technology,Guilin 541006,China;College of Information Science and Engineering,Guilin University of Technology,Guilin 541006,China)

机构地区：[1]桂林理工大学广西嵌入式技术与智能系统重点实验室,广西桂林541006 [2]桂林理工大学信息科学与工程学院,广西桂林541006

出　　处：《计算机工程与设计》2023年第6期1736-1742,共7页Computer Engineering and Design

基　　金：国家自然科学基金项目(61762031);广西科技重大专项基金项目(AA19046004);广西自然科学基金项目(2021JJA170130);广西嵌入式技术与智能系统重点实验室开放基金项目(2020-2-11)。

摘　　要：为解决现有传统环境声音识别技术识别率不高和普通卷积神经网络易出现网络退化的问题,提出一种基于滤波器组和残差网络的环境声音识别算法。采用滤波器组对声音信号进行特征提取,设计14层的残差网络,使用学习率衰减策略,将提取的特征输入到14层残差网络之中训练并测试。实验结果表明,在使用相同数据集ESC-10的情况下,与传统分类器模型和DCASE基线系统提供的识别方法相比,识别准确率分别提高了22.3%、17.4%和9.5%,验证了该方法在小样本情况下具有更高的识别准确率。To solve the problems that the existing traditional environmental sound recognition technology has low recognition rate and the common convolutional neural network is prone to network degradation,an environmental sound recognition method based on a filter bank and residual network was proposed.A filter bank was used to extract features of the sound signal,and a 14-layer residual network was designed,and the learning rate attenuation strategy was used to input the extracted features into the 14-layer residual network for training and testing.Experimental results show that in the case of using the same data set ESC-10,compared with the traditional classifier model and the recognition method provided by the DCASE baseline system,the recognition accuracy is improved by 22.3%,17.4%and 9.5%,respectively.Experimental results verify that in the case of small samples,this method has higher recognition accuracy.

关 键 词：网络退化 滤波器组 残差网络 环境声音识别 特征提取 学习率衰减 分类器模型 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]