Page 163 - 《应用声学》2023年第4期

第 42 卷 第 4 期 Vol. 42, No. 4
2023 年 7 月 Journal of Applied Acoustics July, 2023

⋄ 研究报告 ⋄

基于Mel频谱值和深度学习网络的鸟声识别算法 ∗

李大鹏 周晓彦 † 王基豪 王丽丽 叶 如

(南京信息工程大学电子与信息工程学院 南京 210044)

摘要：为了增强网络对鸟鸣声信号的特征学习能力并提高识别精度，提出一种基于深度残差收缩网络和扩
张卷积的鸟声识别方法。首先，提取鸟鸣声信号的对数 Mel 特征及其一阶和二阶差分系数组成 log-Mel 特征
集，作为网络模型的输入；其次，通过深度残差收缩网络自动学习噪声阈值，减少噪声干扰；然后，引入扩张
卷积增大卷积核感受野并利用注意力机制使网络聚焦于关键帧特征；最后，通过双向长短时记忆网络从学到
的局部特征中学习长期依赖关系。以北京百鸟数据库中的 19 种中国常见鸟类作为实验对象，识别正确率可
以达到 96.58%，并对比模型在不同信噪比数据下的识别结果，结果表明该模型在噪声环境下的识别效果优于
现有模型。
关键词：鸟声识别；log-Mel 特征；深度残差收缩网；扩张卷积神经；注意力机制
中图法分类号: TN912.34 文献标识码: A 文章编号: 1000-310X(2023)04-0825-08
DOI: 10.11684/j.issn.1000-310X.2023.04.018

Bird voice recognition algorithm based on Mel spectrum value and

deep learning network

LI Dapeng ZHOU Xiaoyan WANG Jihao WANG Lili YE Ru

(College of Electronic and Information Engineering, Nanjing University of Information Science and Technology,
Nanjing 210044, China)

Abstract: In order to enhance the feature learning ability of the network to the birdsong signal and improve
the recognition accuracy, a birdsong recognition method based on depth residual shrinkage network (DRSN)
and expanded convolution is proposed. Firstly, the logarithmic Mel feature and its first-order and second-
order difference coefficients of birdsong signal are extracted to form a log-Mel feature set as the input of the
network model; Secondly, the noise threshold is automatically learned through the DRSN to reduce the noise
interference; Then, the expanded convolution is introduced to increase the receptive field of convolution kernel,
and the attention mechanism is used to make the network pay more attention to the characteristics of key
frames; Finally, the long-term dependence is learned from the learned local features through the two-way
long-term and short-term memory network. Taking 19 kinds of common Chinese birds in birdsdata as the
experimental object, the recognition accuracy can reach 96.58%. Compared with the recognition results of the
model under different signal-to-noise ratio data, the results show that the recognition effect of the model in
noise environment is better than that of the existing model.
Keywords: Bird song recognition; log-Mel feature; Depth residual shrinkage network; Dilated convolutions;
Attention mechanism
2022-04-25 收稿; 2022-08-02 定稿
国家自然科学基金项目 (62076064)
∗
作者简介: 李大鹏 (1997– ), 男, 江苏徐州人, 硕士研究生, 研究方向: 模式识别, 声音信号处理。
† 通信作者 E-mail: xiaoyan_zhou@nuist.edu.cn