en
×

分享给微信好友或者朋友圈

使用微信“扫一扫”功能。
作者简介:

吴礼福,男,博士,教授,研究方向为音频信号处理.wulifu@nuist.edu.cn

中图分类号:TN912.3

文献标识码:A

DOI:10.13878/j.cnki.jnuist.20230818001

参考文献 1
张晓宇,仪垂杰.主、被动噪声控制的实验研究[J].噪声与振动控制,2011,31(3):145-148.ZHANG Xiaoyu,YI Chuijie.Experimental study on hybrid noise control[J].Noise and Vibration Control,2011,31(3):145-148
参考文献 2
吴礼福,陈晶晶,郭业才.调节水床效应的双梯度有源噪声控制自适应算法[J].应用声学,2020,39(4):632-637.WU Lifu,CHEN Jingjing,GUO Yecai.Adaptive double-gradient active noise control algorithms for tuning waterbed effect[J].Journal of Applied Acoustics,2020,39(4):632-637
参考文献 3
Schumacher T,Krüger H,Jeub M,et al.Active noise control in headsets:a new approach for broadband feedback ANC[C]//2011 IEEE International Conference on Acoustics,Speech and Signal Processing(ICASSP).May 22-27,2011,Prague,Czech Republic.IEEE,2011:417-420
参考文献 4
Chang C Y,Li S T.Active noise control in headsets by using a low-cost microcontroller[J].IEEE Transactions on Industrial Electronics,2011,58(5):1936-1942
参考文献 5
Bradley S,Backman J,von Hunerbein S,et al.The mechanisms creating wind noise in microphones[C]//Audio Engineering Society Convention 114.Audio Engineering Society,2003:5718
参考文献 6
Mirabilii D,Lodermeyer A,Czwielong F,et al.Simulating wind noise with airflow speed-dependent characteristics[C]//2022 International Workshop on Acoustic Signal Enhancement(IWAENC).September 5-8,2022,Bamberg,Germany.IEEE,2022:1-5
参考文献 7
赖小强,李双田.双传声器系统中的风噪声抑制方法研究[J].信号处理,2013,29(4):436-442.LAI Xiaoqiang,LI Shuangtian.Research for wind noise reduction method in dual-microphone system[J].Journal of Signal Processing,2013,29(4):436-442
参考文献 8
Fisol U M M,Ripin Z M,Ismail N A,et al.Wind noise analysis of a two-way radio[C]//2013 IEEE International Conference on Smart Instrumentation,Measurement and Applications(ICSIMA).November 25-27,2013,Kuala Lumpur,Malaysia.IEEE,2014:1-5
参考文献 9
King B,Atlas L.Coherent modulation comb filtering for enhancing speech in wind noise[C]//International Workshop on Acoustic Echo and Noise Control,2008:14-17
参考文献 10
潘翔.基于风噪声抑制的语音信号增强研究[C]//信号处理在地球物理——浙江省信号处理学会2018年学术年会论文集.杭州,2018:100-105
参考文献 11
Boll S.Suppression of acoustic noise in speech using spectral subtraction[J].IEEE Transactions on Acoustics,Speech,and Signal Processing,1979,27(2):113-120
参考文献 12
Chu Y J,Zhao S P,He L B,et al.Wind noise suppression in filtered-x least mean squares-based active noise control systems[J].The Journal of the Acoustical Society of America,2022,152(6):3340-3345
参考文献 13
吴睿,金向锋.抑制风噪装置的实验[J].应用声学,2022,41(6):1010-1015.WU Rui,JIN Xiangfeng.Experiment on a structure for suppressing the wind noise[J].Journal of Applied Acoustics,2022,41(6):1010-1015
参考文献 14
潘作峰,邓玉伟,郝耀东,等.动态密封状态下汽车风噪性能不确定性研究[J].汽车工程,2021,43(11):1645-1652,1661.PAN Zuofeng,DENG Yuwei,HAO Yaodong,et al.Study on uncertainty of vehicle wind noise performance under dynamic seal condition[J].Automotive Engineering,2021,43(11):1645-1652,1661
参考文献 15
Jiang Y H,Liu H Q,Zhou Y,et al.An integration development of traditional algorithm and neural network for active noise cancellation[C]//2022 IEEE 32nd International Workshop on Machine Learning for Signal Processing(MLSP).August 22-25,2022,Xi'an,China.IEEE,2022:1-6
参考文献 16
Mostafavi A,Cha Y J.Deep learning-based active noise control on construction sites[J].Automation in Construction,2023,151:104885
参考文献 17
Im S,Kim S,Woo S,et al.Deep learning-assisted active noise control in a time-varying environment[J].Journal of Mechanical Science and Technology,2023,37(3):1189-1196
参考文献 18
Cha Y J,Mostafavi A,Benipal S S.DNoiseNet:deep learning-based feedback active noise control in various noisy environments[J].Engineering Applications of Artificial Intelligence,2023,121:105971
参考文献 19
Zhang H,Wang D L.Deep ANC:a deep learning approach to active noise control[J].Neural Networks,2021,141:1-10
参考文献 20
Kwon S,Kim B S,Park J.Active noise reduction with filtered least-mean-square algorithm improved by long short-term memory models for radiation noise of diesel engine[J].Applied Sciences,2022,12(20):10248
参考文献 21
Tan K,Wang D L.Complex spectral mapping with a convolutional recurrent network for monaural speech enhancement[C]//I2019 IEEE International Conference on Acoustics,Speech and Signal Processing(ICASSP).May 12-17,2019,Brighton,UK.IEEE,2019:6865-6869
参考文献 22
郭佳丽,邢双云,栾昊,等.基于改进的LSTM算法的时间序列流量预测[J].南京信息工程大学学报(自然科学版),2021,13(5):571-575.GUO Jiali,XING Shuangyun,LUAN Hao,et al.Prediction of time series traffic based on improved LSTM algorithm[J].Journal of Nanjing University of Information Science & Technology(Natural Science Edition),2021,13(5):571-575
参考文献 23
韩莹,管健,曹允重,等.LSTM-WBLS模型在日降水量预测中的应用[J].南京信息工程大学学报(自然科学版),2023,15(2):180-186.HAN Ying,GUAN Jian,CAO Yunzhong,et al.Application of improved LSTM-WBLS model in daily precipitation forecast[J].Journal of Nanjing University of Information Science & Technology(Natural Science Edition),2023,15(2):180-186
参考文献 24
Park S R,Lee J.A fully convolutional neural network for speech enhancement[J].arXiv e-Print,2016,arXiv:1609.07132
参考文献 25
Varga A,Steeneken H J M.Assessment for automatic speech recognition:Ⅱ.NOISEX-92:a database and an experiment to study the effect of additive noise on speech recognition systems[J].Speech Communication,1993,12(3):247-251
参考文献 26
Nelke C M,Vary P.Measurement,analysis and simulation of wind noise signals for mobile communication devices[C]//2014 14th International Workshop on Acoustic Signal Enhancement(IWAENC).September 8-11,2014,Juan-les-Pins,France.IEEE,2014:327-331
参考文献 27
Allen J B,Berkley D A.Image method for efficiently simulating small-room acoustics[J].The Journal of the Acoustical Society of America,1979,65(4):943-950
参考文献 28
Breining C,Dreiscitel P,Hansler E,et al.Acoustic echo control.An application of very-high-order adaptive filters[J].IEEE Signal Processing Magazine,1999,16(4):42-69
参考文献 29
Reddi S J,Kale S,Kumar S.On the convergence of Adam and beyond[J].arXiv e-Print,2019,arXiv:1904.09237
目录contents

    摘要

    本文提出一种利用双解码卷积循环网络(Dual-decoder Convolutional Recurrent Network,DCRN)代替FxLMS(Filtered-x Least Mean Square)算法的有源噪声控制方法,考虑到相位信息在有源噪声控制(Active Noise Control,ANC)中的重要性,DCRN网络的输入特征为噪声信号的复数频谱(包括实部谱和虚部谱).网络结构中,采用编码模块从噪声复数频谱中提取特征,利用双解码模块分别估计网络输出的实部谱和虚部谱,采用参数共享机制和组策略以降低训练参数的数量并提高网络的学习能力和泛化能力.特别是针对风噪声,选用新的损失函数以及对训练数据进行正则化处理以提升DCRN的性能.实验结果表明,DCRN方法在仿真环境与有源降噪耳机环境下对一般噪声和风噪声都表现出良好的降噪性能和鲁棒性.

    Abstract

    Here,an Active Noise Control (ANC) approach is proposed which replaces Filtered-x Least Mean Square (FxLMS) algorithm with Dual-decoder Convolutional Recurrent Network (DCRN).Due to the importance of phase information in ANC,the input feature of DCRN is the complex spectrogram of the noise signal (including real and imaginary spectrograms).In the network structure,a coding module is used to extract features from the noise complex spectrograms,and a dual-decoder module is used to estimate the real and imaginary spectrograms of the network output.Parameter sharing mechanism and group strategy are adopted to reduce the number of training parameters and improve the learning ability and generalization performance.Especially for wind noise,a new loss function is adopted and the training data are regularized to improve the performance of DCRN.Experiments in both simulation and ANC headphone environments show that the DCRN approach exhibits good noise reduction performance and robustness for both general noise and wind noise.

  • 0 引言

  • 嘈杂环境对人们心理健康和工作造成的负面影响使得噪声控制成为日益关注的焦点.无源噪声控制利用物理材料与噪声声波相互作用消耗声能来衰减大部分中高频噪声[1],但其处理低频噪声的能力有限.有源噪声控制(Active Noise Control,ANC)是一种基于声波干涉原理的噪声控制技术[2],它通过生成与初级噪声幅度相同、相位相反的抵消信号来衰减初级噪声,已有了成功的应用,如有源降噪耳机[3-4]

  • 风噪声作为低频非平稳噪声的典型代表,在实际环境中广泛存在,它是由空气湍流与障碍物(如建筑物、人体、传声器腔体等)之间相互作用产生的[5-7],其能量主要集中在1 000 Hz以下的低频,并且随着频率的增加而滚降[8-9].风噪的非平稳特性使得传统的噪声抑制如基于维纳滤波解的降噪方法表现较差[10-11].虽然有源噪声控制领域已经提出了一些针对风噪声的方法,如Chu等[12]提出基于总最小二乘对滤波x最小均方算法进行修正,提高风噪环境下噪声控制性能,实验结果表明,该算法在有风条件下可以提高约3 dB降噪量.目前大多数针对风噪声的处理方法仍主要基于物理结构来削弱风噪影响,例如:吴睿等[13]基于风噪原理提出一种抗风噪结构抑制风噪对传声器采集信号的影响,该设计由渐变管道和填充泡棉的腔体两部分组成,风噪实验证明了该设计能有效地抑制风噪; 潘作峰等[14]提出一种密封状态下的车内风噪性能区间不确定性的分析和优化方法,该方法可在保证相关零部件质量基本不变的前提下,降低车内噪声水平及其波动幅度,明显提升系统的稳健性.

  • 近年来,神经网络在有源噪声控制中的运用引起了关注.Jiang等[15]首先利用经典ANC算法处理噪声,然后将残余噪声和估计噪声送入神经网络以消除非线性失真.Mostafavi等[16]提出一种基于深度学习的前馈ANC控制器,通过考虑声学设备的延迟和非线性以衰减与施工相关的噪声.Im等[17]提出一种深度学习辅助的次级路径更新技术,训练深度神经网络根据不断变化的边界条件实时估计次级路径以应对时变环境.Cha等[18]提出一种基于DNoiseNet的反馈有源降噪方法,通过在各种环境中学习不同噪声的多级时间特征,克服了传统有源降噪的局限性,在此基础上,还提出了基于多层感知机(Multilayer Perception,MLP)神经网络的次级路径估计器以提高DNoiseNet的性能.Zhang等[19]提出一种Deep-ANC方法以解决非线性ANC问题,该方法训练网络以编码不同噪声环境下的最优控制参数,还引入延迟补偿策略解决ANC系统潜在的延迟问题.Kwon等[20]采用长短期记忆(Long Short-Term Memory,LSTM)层作为ANC控制器,生成控制信号以最小化残留噪声.

  • 鉴于风噪声的非平稳、非线性特性和神经网络强大的非线性映射能力,本文研究了双解码卷积循环网络(Dual-decoder Convolutional Recurrent Network,DCRN)[21]代替传统FxLMS(Filtered-x Least Mean Square)算法的ANC方法.考虑到相位信息在ANC中的重要性,DCRN网络的输入特征为噪声信号的复数频谱(包括实部谱和虚部谱).为了降低训练参数的数量并提高网络的学习能力和泛化能力,网络结构中采用了参数共享机制和组策略.

  • 1 算法描述

  • 1.1 FxLMS算法

  • 因其简单、稳定和相对低的计算复杂度,FxLMS及其扩展算法在ANC中应用广泛.FxLMS算法是以LMS算法为基础,在参考信号与滤波器之间放置估计的次级路径以补偿次级路径对有源噪声控制系统的影响.单通道前馈FxLMS算法框图如图1所示.

  • 图1 单通道前馈FxLMS算法框图

  • Fig.1 Block diagram of single-channel feed-forward FxLMS

  • 参考信号xn)由参考麦克风拾取并通过自适应滤波器Wz)生成次级信号yn),如式(1).参考信号和次级信号分别经过初级路径Pz)和次级路径Sz)得到期望信号dn)和抵消信号ysn),噪声残留由误差麦克风拾取得到.FxLMS算法对滤波器Wz)进行自适应迭代,生成抵消信号ysn)以最小化噪声残留信号en),如式(2).Wz)迭代更新方程如式(3).

  • y(n)=wT(n)x(n),
    (1)
  • e(n)=d(n)-ys(n),
    (2)
  • w(n+1)=w(n)+μe(n)[x(n)*s^(n)].
    (3)
  • 式(3)中:wn)为自适应滤波器权重; μ是自适应步长参数; s^n为次级路径估计s^z的脉冲响应; *表示线性卷积运算.

  • 1.2 DCRN方法

  • 1.2.1 算法原理

  • 图2所示为单通道前馈DCRN方法的系统框图.相较于图1,DCRN方法采用DCRN代替Wz).参考信号xn)经过短时傅里叶变换(Short-time Fourier Transform,STFT)后得到复频谱,其实部和虚部频谱分别为Xrtf)和Xitf),其中,tf表示为时频(T-F)单元内的时间和频率,下标ri分别表示信号频谱的实部和虚部.将Xrtf)和Xitf)作为输入送至DCRN并计算出次级信号的实部与虚部频谱即Yrtf)和Yitf).次级信号的复频谱经逆短时傅里叶变换得到时域信号yn),与次级路径Sz)卷积得到抵消信号ysn).dn)和ysn)用于计算损失函数,训练DCRN参数以最小化误差信号en).

  • 图2 单通道前馈DCRN方法系统框图

  • Fig.2 Block diagram of single-channel feed-forward DCRN

  • 1.2.2 网络参数

  • 图2中DCRN具体结构如图3所示.DCRN结构包括3个部分:编码模块、时序建模模块和解码模块.编码模块和解码模块分别包含5层卷积层和5层反卷积层,位于编码模块和解码模块之间的时序建模模块由采用组策略的2个长短期记忆层(LSTM)组成[22-23],其中组数设置为2.采用组策略的LSTM层有助于提高循环层的效率和降低模型的复杂度.编码模块从输入信号复频谱中提取出高维特征,而解码模块以逆向操作将编码模块提取的高维特征逐层恢复至目标信号.

  • 图3中Xrtf)和Xitf)被视为2个不同的输入通道,编码模块和时序建模模块共享Xrtf)和Xitf),而2个解码模块分别用于估计DCRN网络输出信号的实部和虚部频谱.这样的设置使得DCRN可以同时处理实部和虚部的复频谱映射,将估计实部和虚部视为2个相关子任务,从而通过参数共享鼓励学习和提高泛化能力.在模型训练过程中,采用了跳跃连接方法(skip connection)[24],该方法将编码模块每个卷积层输出与双解码模块中对应的反卷积层的输入相连接,可以逐层传递编码模块中提取的特征到解码模块,避免输入信息在传递过程中丢失过多,从而改善通过DCRN的信息和梯度.DCRN网络在除输出层外的所有卷积层和反卷积层中均使用指数线性单元(Exponential Linear Units,ELU)作为激活函数,并在每次卷积(或反卷积)之后采用批量归一化操作(Batch Normalization,BN)来加快训练速度和提高模型稳定性.此外,在2个解码模块的输出层之后使用线性层来预测输出信号的实部和虚部频谱.

  • 表1给出了DCRN模型每层网络的具体参数设置,其中,T表示STFT幅度谱中时间帧的数量,双解码模块的反卷积层参数相同.表1中:卷积层和反卷积层的输入维度和输出维度括号内3个参数分别表示特征通道数、时间总帧数、特征维度,卷积层和反卷积层的层间超参数分别对应卷积核、卷积步幅、输出通道数; 5层卷积层的输出通道数分别为8、16、32、64、128,5层反卷积层的输出通道数分别为64、32、16、8、1; 卷积层和反卷积层的时间维度和特征维度的卷积步幅设置为(1,2),卷积核均设置为(1,3).此外,为了聚合频谱的上下文,所有卷积层和反卷积层中沿频率维度采用的步长为2.值得注意的是,由于采用了跳跃连接方法,解码模块中每层反卷积层的输入特征通道数是对应的卷积层输入特征通道数2倍.这样的设计使得编码模块提取的特征能够更加有效地传递到解码模块,从而有助于保持和恢复输入信号的重要特征.

  • 图3 DCRN的具体结构

  • Fig.3 Specific structure of DCRN

  • 表1 DCRN参数

  • Table1 DCRN parameters

  • 1.2.3 特征提取和损失函数

  • 网络输入参考信号xn)均以16 kHz采样,使用20 ms的汉明窗将参考信号xn)分割成一系列时间帧,相邻两帧之间的重叠为10 ms.对每帧应用320点的短时傅里叶变换(STFT)生成161维的实部和虚部频谱.

  • 本文选择复频谱映射是由于相位在有源噪声控制中的重要性.ANC的目标是产生抵消信号以衰减误差麦克风处的初级噪声,均方误差(Mean Squared Error,MSE)损失函数是最广泛使用的处理回归问题的损失函数之一.式(4)为MSE损失函数的计算公式,式中,N为样本数量即每次迭代数据中批处理量大小,yk,trueyk,predicted分别为第k个样本的真实值和对应的预测值,在本文中分别是dn)与ysn)经STFT变换后的复频谱.

  • eMSE=1NΣk=0Nyk, predicted -yk, true 2
    (4)
  • 2 实验结果与分析

  • 2.1 实验配置

  • 实验分为两部分,分别验证DCRN方法在一般噪声和风噪声条件下的降噪性能.第一部分实验使用Sound-Effect音效库General Series 6000系列(https://soundeffects.fandom.com/wiki/The_General_Series_6000_Sound_Effects_Library)中多种噪声信号创建训练集.验证集采用NOISEX-92数据集[25]中的Leopard噪声,测试集使用NOISEX-92中的babble噪声、engine噪声和factory噪声.第二部分实验针对风噪声进行训练,由于目前缺乏公开的风噪声数据集,本文采用Nelke等[26]提出的合成风噪声信号模型(后文称“Nelke模型”)生成风噪声信号,包括阵风和恒定风两种类型.风噪声的数据集设置为:Nelke模型生成约10 h的恒定风噪声作为训练集,约1 h的阵风风噪声作为验证集,在户外场景采集的正常风速和强风速两种条件下真实风噪声(https://www.iks.rwth-aachen.de/en/research/tools-downloads/databases/wind-noise-database/)作为测试集.

  • 在每一部分实验中,为了验证网络的泛化性能,采用两类初级路径和次级路径进行实验.第一类采用镜像源法[27]生成房间冲击响应(Room Impulse Responses,RIRs).本文模拟了一个3 m×4 m×2 m的房间,其中,参考麦克风位于(1.5,1,1)m处,次级源(抵消扬声器)位于(1.5,2.5,1)m处,误差麦克风位于(1.5,3,1)m处.模拟生成的RIRs混响时间T60∈{0.15,0.175,0.2,0.225,0.25}s[28],阶数为512.每种混响时间生成两个RIRs,分别用作初级路径和次级路径.在训练阶段随机选择混响时间T60∈{0.15,0.175,0.225,0.25}s中的一组RIRs,测试阶段采用混响时间为0.2 s的RIRs作为默认测试RIRs.第二类旨在验证算法在有源降噪耳机中的实际效果,其初级路径与次级路径来自于吴礼福等[2]实测的有源降噪耳机场景,阶数分别为512和256.

  • 对于两部分实验,分别创建了20 000条训练数据、100条验证数据和100条测试数据.每条训练数据、验证数据以及测试数据均是从对应数据集中随机裁剪的6 s信号创建.在训练模型时使用AMSGard优化器[29]更新网络权重,初始学习率设置为0.001,采用StepLR机制在训练过程中动态调整学习率,批大小设为16,训练总轮数设为30.

  • 2.2 实验结果分析

  • 本文以降噪量评价算法的降噪性能,其中,降噪量定义为算法作用前后残留噪声的功率谱密度差异.对比算法选择FxLMS算法与文献[19]中提出的Deep-ANC方法.FxLMS算法滤波器长度设为512; Deep-ANC方法使用与DCRN方法相同的超参数和数据集.

  • 式(4)中的平方运算会放大较大误差的影响(对异常值比较敏感),导致模型对异常值过度拟合.平均绝对值误差(Mean Absolute Error,MAE)损失函数由于使用绝对值计算使其对异常值的鲁棒性更好.由于风噪声信号呈现能量集中在低频部分并且随着频率升高能量滚降,而MAE对所有误差赋予相同权重,可以更平衡地考虑低频与高频之间的差异.因此,在风噪声中优先采用MAE损失函数,同时将MAE损失函数用于一般噪声中进行对比.式(5)为MAE损失函数计算公式,式中各参数同式(4).后文中对使用MSE和MAE损失函数的DCRN方法分别命名为DCRN-MSE方法和DCRN-MAE方法.

  • eMAE=1NΣk=0Nyk, predicted -yk, true .
    (5)
  • 2.2.1 一般噪声

  • 实验1采用镜像源法生成的RIRs作为初级路径和次级路径以验证DCRN在仿真环境中降噪性能.选择4种噪声情况进行测试,分别为babble、engine、babble转变为engine以及babble混合factory.图4a和4b分别给出了babble和engine两种噪声下FxLMS、Deep-ANC和DCRN方法降噪后误差信号的功率谱对比.观察图4可以发现DCRN方法和Deep-ANC方法在babble和engine两种噪声下相较于FxLMS算法均表现出更好的降噪性能.传统的FxLMS算法降噪性能仅限于低频窄带部分,而DCRN方法在宽带和高频部分均表现出更高的降噪性能.

  • 实际环境中可能存在噪声类型转换或多种噪声同时存在的情况,对babble转变为engine以及babble混合factory两种噪声情况进行测试,以验证DCRN方法的鲁棒性.图5a和5b分别给出了babble转变为engine噪声与babble混合factory噪声两种情况下,FxLMS、Deep-ANC和DCRN方法降噪后误差信号的功率谱对比.观察图5可以发现,在噪声类型转变以及多种噪声混合的情况下,DCRN方法相较于FxLMS算法在宽频带内降噪性能均有明显提升.实验1的结果表明,在仿真环境下DCRN方法在宽带噪声上均能够实现更好的降噪性能,并且在应对噪声变化的情况下表现出良好的鲁棒性.对比DCRN-MAE方法与DCRN-MSE方法可以发现在低频部分前者降噪量低于后者,表明在仿真环境中DCRN-MAE方法降噪性能略差于DCRN-MSE方法.

  • 图4 babble与engine两种噪声降噪前后误差信号功率谱

  • Fig.4 Error signal power spectrums before and after ANC a.babble noise; b.engine noise

  • 图5 babble转变为engine与babble混合factory两种情况降噪前后误差信号功率谱

  • Fig.5 Error signal power spectrums before and after ANC a.from babble noise to engine noise; b.babble noise mixed with factory noise

  • 为了进一步验证DCRN方法在实际环境下的降噪性能,实验2选择了有源降噪耳机场景进行验证.图6a和6b分别给出了babble和factory两种噪声下,FxLMS、Deep-ANC和DCRN方法降噪后误差信号的功率谱对比.可以看出,在有源降噪耳机环境下,DCRN方法在2 000 Hz以内的低频部分相较于FxLMS算法降噪量提升5~10 dB,在整个宽频带内降噪性能略优于Deep-ANC方法.图7a和7b分别给出了babble转变为engine噪声和babble混合factory噪声两种情况下,FxLMS、Deep-ANC和DCRN方法降噪后误差信号的功率谱对比.可以发现在有源降噪耳机环境下,对于噪声类型转变和噪声混合的情况,DCRN方法在2 000 Hz以内的低频部分相较于FxLMS算法降噪量均能提升5~10 dB.实验2的结果表明,对于一般噪声,在有源降噪耳机环境中DCRN方法在低频段相较于FxLMS算法和Deep-ANC方法均有降噪量的提升,并且在应对噪声变化时表现出了较好的鲁棒性.对比DCRN-MAE方法与DCRN-MSE方法可以发现在低频段前者降噪量较于后者提升5~10 dB,并且解决了后者在高频部分能量不降反增的问题.该结果表明,在有源降噪耳机环境中对于一般噪声DCRN-MAE方法降噪性能优于DCRN-MSE方法.

  • 2.2.2 风噪声

  • 实验3采用模拟生成的房间脉冲响应(RIRs)作为初级路径和次级路径,验证在仿真环境下,DCRN方法降低风噪的性能.选择户外采集的正常风速和强风速两种风噪声进行测试.图8a和8b分别给出了正常风速和强风速两种风噪下,FxLMS、Deep-ANC和两种DCRN方法降噪后误差信号的功率谱对比,可以发现风噪声能量主要集中的1 000 Hz以内,在正常风速和强风速两种风噪下两种DCRN方法相较于FxLMS算法降噪量均提高10 dB左右,DCRN-MAE方法在整体降低风噪的能力上明显优于Deep-ANC和DCRN-MSE方法.实验3验证了DCRN方法在风噪声环境中降低风噪声的可行性以及DCRN-MAE方法降低风噪的性能优于DCRN-MSE方法.

  • 图6 有源降噪耳机环境下babble与factory两种噪声降噪前后误差信号功率谱

  • Fig.6 Error signal power spectrums before and after ANC in active headphone environment a.babble noise; b.factory noise

  • 图7 有源降噪耳机环境下babble变为engine与babble混合factory两种情况降噪前后误差信号功率谱

  • Fig.7 Error signal power spectrums before and after ANC in active headphone environment a.from babble noise to engine noise; b.babble noise mixed with factory noise

  • 图8 仿真环境下正常风速和强风速两种风噪降噪前后误差信号功率谱

  • Fig.8 Error signal power spectrums before and after ANC in simulated environment a.normal speed wind noise; b.strong speed wind noise

  • 从图8中可以观察到DCRN-MSE和Deep-ANC方法在风噪声的高频部分(1 000 Hz以上)表现较差.这是由于风噪信号在高频部分的能量滚降迅速,其能量较低频而言非常小,导致DCRN-MSE和Deep-ANC方法在训练模型时对风噪高频部分出现训练过拟合.对此可以通过正则化即提高风噪声信号高频部分的能量来解决模型出现过拟合的现象.考虑到风噪的特点,本文在风噪信号1 000 Hz频率以上叠加白噪声进行正则化处理.图9是风噪信号高频部分正则化处理前后的功率谱图.从图9中可以观察到正则化处理不改变风噪信号的主要结构,仅处理风噪信号高频部分不影响风噪能量主要集中的低频段.

  • 实验4验证了DCRN方法在实际环境即有源降噪耳机中降低风噪的性能.风噪数据集的使用同实验3一致,不同的是对风噪数据集进行正则化处理.图10a和10b分别是在正常风速和强风速两种风噪声情况下FxLMS、Deep-ANC和两种DCRN方法降噪后误差信号的功率谱对比.可以发现在风噪声能量主要集中的1 000 Hz以内的低频段,DCRN-MSE和Deep-ANC方法对于两种风速风噪情况仅在600 Hz以内相较于FxLMS算法有降噪量的提升,DCRN-MAE方法在1 000 Hz以内相较于DCRN-MSE方法降噪量均有显著提升.实验4结果表明,采用正则化训练数据后,在有源降噪耳机环境下DCRN-MAE方法在降低风噪性能上相较于DCRN-MSE方法具有明显优势.

  • 图9 风噪信号正则化处理前后功率谱

  • Fig.9 Power spectrums of wind noise signal before and after regularization

  • 图10 有源降噪耳机环境下正常风速和强风速风噪降噪前后误差信号功率谱

  • Fig.10 Error signal power spectrums before and after ANC in active headphone environment a.normal speed wind noise; b.strong speed wind noise

  • 实验5对比采用噪声信号目标幅度谱(Target Magnitude Spectrum,TMS)作为输入特征的CRN以验证融合相位估计的复频谱映射的优势.幅度谱映射的CRN具体结构如图11所示,相较于DCRN采用双解码模块分别估计网络输出信号的实部谱和虚部谱,幅度谱映射的CRN仅需一个解码模块即可.CRN中编解码模块和时序模块中各层网络具体参数与DCRN相同.本文将采用幅度谱映射的CRN方法简称为CRN-TMS方法,训练CRN的数据集同实验4且采用MAE损失函数.图12a和12b分别是在正常风速和强风速两种风噪声情况下FxLMS、Deep ANC、DCRN-MAE方法和CRN-TMS方法降噪后误差信号的功率谱对比.可以观察到采用幅度谱映射的CRN-TMS方法降低风噪性能弱于传统FxLMS算法,表明ANC系统中相位信息的重要性,也表明采用复频谱的DCRN-MAE相较于仅采用幅度谱的CRN-TMS在融合相位信息后降低风噪具有明显的优势.

  • 算法的计算复杂度在评估算法运行时消耗的计算资源量方面起主要作用.深度学习算法一般采用浮点数运算次数(Floating Point Operations,FLOPs)(包括乘法运算次数和加法运算次数)衡量模型计算复杂度.本文采用广泛使用的Python中thop库(https://pypi.org/project/thop/)计算DCRN的FLOPs和训练参数量.在计算复杂度方面,DCRN模型中可训练参数约为2.32×106,每个时间帧的浮点运算次数为4.13×106.对比算法Deep-ANC模型的可训练参数和浮点运算次数分别为8.83×106和12.78×106,DCRN模型相较于Deep-ANC模型大幅降低了网络的可训练参数和计算复杂度.在同一硬件平台上,传统FxLMS算法、DCRN方法与Deep-ANC方法对于1 s数据的平均处理时间(运行100次)如表2所示,DCRN方法与Deep-ANC方法比传统自适应滤波算法更快.这是由于深度学习网络训练完成后,其网络权重是固定的,因此可以利用硬件平台进行并行运算,从而减少运行时间.

  • 图11 幅度谱映射的CRN具体结构

  • Fig.11 Specific structure of CRN for magnitude spectral mapping

  • 图12 有源降噪耳机环境下正常风速和强风速风噪降噪前后误差信号功率谱

  • Fig.12 Error signal power spectrums before and after ANC in active headphone environment a.normal speed wind noise; b.strong speed wind noise

  • 表2 不同算法处理1 s数据的平均运行时间

  • Table2 Average execution time for processing1 second of data by different algorithms

  • 3 结论

  • 本文研究了基于双解码卷积循环网络的DCRN有源噪声控制方法,利用DCRN代替自适应滤波器来编码,适用于不同噪声环境下ANC控制器最优控制参数,通过训练DCRN分别估计与参考信号对应的次级信号的实部和虚部,进而使抵消信号能够最大化衰减误差麦克风处的噪声残留.此外,本文采用复数频谱作为输入特征和训练目标,充分考虑了相位在有源噪声控制中的重要性.针对一般噪声与风噪声分别在仿真环境和有源降噪耳机环境下进行实验.实验结果表明,对于一般噪声DCRN方法具有良好的降噪性能和鲁棒性.在风噪声中,正则化训练数据后,采用MAE损失函数的DCRN方法相比于采用MSE损失函数的DCRN方法降噪风噪的性能明显提升.实验结果说明了损失函数对网络性能的显著影响.未来的研究将进一步探索采用不同形式的损失函数如基于对抗网络的损失函数,以尝试进一步提高网络性能.

  • 参考文献

    • [1] 张晓宇,仪垂杰.主、被动噪声控制的实验研究[J].噪声与振动控制,2011,31(3):145-148.ZHANG Xiaoyu,YI Chuijie.Experimental study on hybrid noise control[J].Noise and Vibration Control,2011,31(3):145-148

    • [2] 吴礼福,陈晶晶,郭业才.调节水床效应的双梯度有源噪声控制自适应算法[J].应用声学,2020,39(4):632-637.WU Lifu,CHEN Jingjing,GUO Yecai.Adaptive double-gradient active noise control algorithms for tuning waterbed effect[J].Journal of Applied Acoustics,2020,39(4):632-637

    • [3] Schumacher T,Krüger H,Jeub M,et al.Active noise control in headsets:a new approach for broadband feedback ANC[C]//2011 IEEE International Conference on Acoustics,Speech and Signal Processing(ICASSP).May 22-27,2011,Prague,Czech Republic.IEEE,2011:417-420

    • [4] Chang C Y,Li S T.Active noise control in headsets by using a low-cost microcontroller[J].IEEE Transactions on Industrial Electronics,2011,58(5):1936-1942

    • [5] Bradley S,Backman J,von Hunerbein S,et al.The mechanisms creating wind noise in microphones[C]//Audio Engineering Society Convention 114.Audio Engineering Society,2003:5718

    • [6] Mirabilii D,Lodermeyer A,Czwielong F,et al.Simulating wind noise with airflow speed-dependent characteristics[C]//2022 International Workshop on Acoustic Signal Enhancement(IWAENC).September 5-8,2022,Bamberg,Germany.IEEE,2022:1-5

    • [7] 赖小强,李双田.双传声器系统中的风噪声抑制方法研究[J].信号处理,2013,29(4):436-442.LAI Xiaoqiang,LI Shuangtian.Research for wind noise reduction method in dual-microphone system[J].Journal of Signal Processing,2013,29(4):436-442

    • [8] Fisol U M M,Ripin Z M,Ismail N A,et al.Wind noise analysis of a two-way radio[C]//2013 IEEE International Conference on Smart Instrumentation,Measurement and Applications(ICSIMA).November 25-27,2013,Kuala Lumpur,Malaysia.IEEE,2014:1-5

    • [9] King B,Atlas L.Coherent modulation comb filtering for enhancing speech in wind noise[C]//International Workshop on Acoustic Echo and Noise Control,2008:14-17

    • [10] 潘翔.基于风噪声抑制的语音信号增强研究[C]//信号处理在地球物理——浙江省信号处理学会2018年学术年会论文集.杭州,2018:100-105

    • [11] Boll S.Suppression of acoustic noise in speech using spectral subtraction[J].IEEE Transactions on Acoustics,Speech,and Signal Processing,1979,27(2):113-120

    • [12] Chu Y J,Zhao S P,He L B,et al.Wind noise suppression in filtered-x least mean squares-based active noise control systems[J].The Journal of the Acoustical Society of America,2022,152(6):3340-3345

    • [13] 吴睿,金向锋.抑制风噪装置的实验[J].应用声学,2022,41(6):1010-1015.WU Rui,JIN Xiangfeng.Experiment on a structure for suppressing the wind noise[J].Journal of Applied Acoustics,2022,41(6):1010-1015

    • [14] 潘作峰,邓玉伟,郝耀东,等.动态密封状态下汽车风噪性能不确定性研究[J].汽车工程,2021,43(11):1645-1652,1661.PAN Zuofeng,DENG Yuwei,HAO Yaodong,et al.Study on uncertainty of vehicle wind noise performance under dynamic seal condition[J].Automotive Engineering,2021,43(11):1645-1652,1661

    • [15] Jiang Y H,Liu H Q,Zhou Y,et al.An integration development of traditional algorithm and neural network for active noise cancellation[C]//2022 IEEE 32nd International Workshop on Machine Learning for Signal Processing(MLSP).August 22-25,2022,Xi'an,China.IEEE,2022:1-6

    • [16] Mostafavi A,Cha Y J.Deep learning-based active noise control on construction sites[J].Automation in Construction,2023,151:104885

    • [17] Im S,Kim S,Woo S,et al.Deep learning-assisted active noise control in a time-varying environment[J].Journal of Mechanical Science and Technology,2023,37(3):1189-1196

    • [18] Cha Y J,Mostafavi A,Benipal S S.DNoiseNet:deep learning-based feedback active noise control in various noisy environments[J].Engineering Applications of Artificial Intelligence,2023,121:105971

    • [19] Zhang H,Wang D L.Deep ANC:a deep learning approach to active noise control[J].Neural Networks,2021,141:1-10

    • [20] Kwon S,Kim B S,Park J.Active noise reduction with filtered least-mean-square algorithm improved by long short-term memory models for radiation noise of diesel engine[J].Applied Sciences,2022,12(20):10248

    • [21] Tan K,Wang D L.Complex spectral mapping with a convolutional recurrent network for monaural speech enhancement[C]//I2019 IEEE International Conference on Acoustics,Speech and Signal Processing(ICASSP).May 12-17,2019,Brighton,UK.IEEE,2019:6865-6869

    • [22] 郭佳丽,邢双云,栾昊,等.基于改进的LSTM算法的时间序列流量预测[J].南京信息工程大学学报(自然科学版),2021,13(5):571-575.GUO Jiali,XING Shuangyun,LUAN Hao,et al.Prediction of time series traffic based on improved LSTM algorithm[J].Journal of Nanjing University of Information Science & Technology(Natural Science Edition),2021,13(5):571-575

    • [23] 韩莹,管健,曹允重,等.LSTM-WBLS模型在日降水量预测中的应用[J].南京信息工程大学学报(自然科学版),2023,15(2):180-186.HAN Ying,GUAN Jian,CAO Yunzhong,et al.Application of improved LSTM-WBLS model in daily precipitation forecast[J].Journal of Nanjing University of Information Science & Technology(Natural Science Edition),2023,15(2):180-186

    • [24] Park S R,Lee J.A fully convolutional neural network for speech enhancement[J].arXiv e-Print,2016,arXiv:1609.07132

    • [25] Varga A,Steeneken H J M.Assessment for automatic speech recognition:Ⅱ.NOISEX-92:a database and an experiment to study the effect of additive noise on speech recognition systems[J].Speech Communication,1993,12(3):247-251

    • [26] Nelke C M,Vary P.Measurement,analysis and simulation of wind noise signals for mobile communication devices[C]//2014 14th International Workshop on Acoustic Signal Enhancement(IWAENC).September 8-11,2014,Juan-les-Pins,France.IEEE,2014:327-331

    • [27] Allen J B,Berkley D A.Image method for efficiently simulating small-room acoustics[J].The Journal of the Acoustical Society of America,1979,65(4):943-950

    • [28] Breining C,Dreiscitel P,Hansler E,et al.Acoustic echo control.An application of very-high-order adaptive filters[J].IEEE Signal Processing Magazine,1999,16(4):42-69

    • [29] Reddi S J,Kale S,Kumar S.On the convergence of Adam and beyond[J].arXiv e-Print,2019,arXiv:1904.09237

  • 参考文献

    • [1] 张晓宇,仪垂杰.主、被动噪声控制的实验研究[J].噪声与振动控制,2011,31(3):145-148.ZHANG Xiaoyu,YI Chuijie.Experimental study on hybrid noise control[J].Noise and Vibration Control,2011,31(3):145-148

    • [2] 吴礼福,陈晶晶,郭业才.调节水床效应的双梯度有源噪声控制自适应算法[J].应用声学,2020,39(4):632-637.WU Lifu,CHEN Jingjing,GUO Yecai.Adaptive double-gradient active noise control algorithms for tuning waterbed effect[J].Journal of Applied Acoustics,2020,39(4):632-637

    • [3] Schumacher T,Krüger H,Jeub M,et al.Active noise control in headsets:a new approach for broadband feedback ANC[C]//2011 IEEE International Conference on Acoustics,Speech and Signal Processing(ICASSP).May 22-27,2011,Prague,Czech Republic.IEEE,2011:417-420

    • [4] Chang C Y,Li S T.Active noise control in headsets by using a low-cost microcontroller[J].IEEE Transactions on Industrial Electronics,2011,58(5):1936-1942

    • [5] Bradley S,Backman J,von Hunerbein S,et al.The mechanisms creating wind noise in microphones[C]//Audio Engineering Society Convention 114.Audio Engineering Society,2003:5718

    • [6] Mirabilii D,Lodermeyer A,Czwielong F,et al.Simulating wind noise with airflow speed-dependent characteristics[C]//2022 International Workshop on Acoustic Signal Enhancement(IWAENC).September 5-8,2022,Bamberg,Germany.IEEE,2022:1-5

    • [7] 赖小强,李双田.双传声器系统中的风噪声抑制方法研究[J].信号处理,2013,29(4):436-442.LAI Xiaoqiang,LI Shuangtian.Research for wind noise reduction method in dual-microphone system[J].Journal of Signal Processing,2013,29(4):436-442

    • [8] Fisol U M M,Ripin Z M,Ismail N A,et al.Wind noise analysis of a two-way radio[C]//2013 IEEE International Conference on Smart Instrumentation,Measurement and Applications(ICSIMA).November 25-27,2013,Kuala Lumpur,Malaysia.IEEE,2014:1-5

    • [9] King B,Atlas L.Coherent modulation comb filtering for enhancing speech in wind noise[C]//International Workshop on Acoustic Echo and Noise Control,2008:14-17

    • [10] 潘翔.基于风噪声抑制的语音信号增强研究[C]//信号处理在地球物理——浙江省信号处理学会2018年学术年会论文集.杭州,2018:100-105

    • [11] Boll S.Suppression of acoustic noise in speech using spectral subtraction[J].IEEE Transactions on Acoustics,Speech,and Signal Processing,1979,27(2):113-120

    • [12] Chu Y J,Zhao S P,He L B,et al.Wind noise suppression in filtered-x least mean squares-based active noise control systems[J].The Journal of the Acoustical Society of America,2022,152(6):3340-3345

    • [13] 吴睿,金向锋.抑制风噪装置的实验[J].应用声学,2022,41(6):1010-1015.WU Rui,JIN Xiangfeng.Experiment on a structure for suppressing the wind noise[J].Journal of Applied Acoustics,2022,41(6):1010-1015

    • [14] 潘作峰,邓玉伟,郝耀东,等.动态密封状态下汽车风噪性能不确定性研究[J].汽车工程,2021,43(11):1645-1652,1661.PAN Zuofeng,DENG Yuwei,HAO Yaodong,et al.Study on uncertainty of vehicle wind noise performance under dynamic seal condition[J].Automotive Engineering,2021,43(11):1645-1652,1661

    • [15] Jiang Y H,Liu H Q,Zhou Y,et al.An integration development of traditional algorithm and neural network for active noise cancellation[C]//2022 IEEE 32nd International Workshop on Machine Learning for Signal Processing(MLSP).August 22-25,2022,Xi'an,China.IEEE,2022:1-6

    • [16] Mostafavi A,Cha Y J.Deep learning-based active noise control on construction sites[J].Automation in Construction,2023,151:104885

    • [17] Im S,Kim S,Woo S,et al.Deep learning-assisted active noise control in a time-varying environment[J].Journal of Mechanical Science and Technology,2023,37(3):1189-1196

    • [18] Cha Y J,Mostafavi A,Benipal S S.DNoiseNet:deep learning-based feedback active noise control in various noisy environments[J].Engineering Applications of Artificial Intelligence,2023,121:105971

    • [19] Zhang H,Wang D L.Deep ANC:a deep learning approach to active noise control[J].Neural Networks,2021,141:1-10

    • [20] Kwon S,Kim B S,Park J.Active noise reduction with filtered least-mean-square algorithm improved by long short-term memory models for radiation noise of diesel engine[J].Applied Sciences,2022,12(20):10248

    • [21] Tan K,Wang D L.Complex spectral mapping with a convolutional recurrent network for monaural speech enhancement[C]//I2019 IEEE International Conference on Acoustics,Speech and Signal Processing(ICASSP).May 12-17,2019,Brighton,UK.IEEE,2019:6865-6869

    • [22] 郭佳丽,邢双云,栾昊,等.基于改进的LSTM算法的时间序列流量预测[J].南京信息工程大学学报(自然科学版),2021,13(5):571-575.GUO Jiali,XING Shuangyun,LUAN Hao,et al.Prediction of time series traffic based on improved LSTM algorithm[J].Journal of Nanjing University of Information Science & Technology(Natural Science Edition),2021,13(5):571-575

    • [23] 韩莹,管健,曹允重,等.LSTM-WBLS模型在日降水量预测中的应用[J].南京信息工程大学学报(自然科学版),2023,15(2):180-186.HAN Ying,GUAN Jian,CAO Yunzhong,et al.Application of improved LSTM-WBLS model in daily precipitation forecast[J].Journal of Nanjing University of Information Science & Technology(Natural Science Edition),2023,15(2):180-186

    • [24] Park S R,Lee J.A fully convolutional neural network for speech enhancement[J].arXiv e-Print,2016,arXiv:1609.07132

    • [25] Varga A,Steeneken H J M.Assessment for automatic speech recognition:Ⅱ.NOISEX-92:a database and an experiment to study the effect of additive noise on speech recognition systems[J].Speech Communication,1993,12(3):247-251

    • [26] Nelke C M,Vary P.Measurement,analysis and simulation of wind noise signals for mobile communication devices[C]//2014 14th International Workshop on Acoustic Signal Enhancement(IWAENC).September 8-11,2014,Juan-les-Pins,France.IEEE,2014:327-331

    • [27] Allen J B,Berkley D A.Image method for efficiently simulating small-room acoustics[J].The Journal of the Acoustical Society of America,1979,65(4):943-950

    • [28] Breining C,Dreiscitel P,Hansler E,et al.Acoustic echo control.An application of very-high-order adaptive filters[J].IEEE Signal Processing Magazine,1999,16(4):42-69

    • [29] Reddi S J,Kale S,Kumar S.On the convergence of Adam and beyond[J].arXiv e-Print,2019,arXiv:1904.09237

  • Address:No. 219, Ningliu Road, Nanjing, Jiangsu Province

    Postcode:210044

    Phone:025-58731025