结合LSTM与ResNet的声学回声消除  

Acoustic echo cancellation by combining LSTM and ResNet

作　　者：许春冬[1] 徐锦武 王茹霞 凌贤鹏 黄乔月 郭桥生 XU Chundong;XU Jinwu;WANG Ruxia;LING Xianpeng;HUANG Qiaoyue;GUO Qiaosheng(School of Information Engineering,Jiangxi University of Science and Technology,Ganzhou 341000,China;Anker Innovations Technology Co Ltd,Shenzhen 518000,China;Zhaoyang Gevotai(Xinfeng)Technology Co Ltd,Ganzhou 341600,China)

机构地区：[1]江西理工大学信息工程学院,江西赣州341000 [2]安克创新科技股份有限公司,广东深圳518000 [3]朝阳聚声泰(信丰)科技有限公司,江西赣州341600

出　　处：《传感器与微系统》2023年第5期29-32,共4页Transducer and Microsystem Technologies

基　　金：国家自然科学基金资助项目(11864016,61671442)。

摘　　要：针对传统的声学回声消除(AEC)方法在双端讲话场景下较难实现快速收敛和动态自适应的问题,提出了一种结合长短时记忆(LSTM)与残差神经网络(ResNet)的AEC方法。通过使用LSTM和ResNet相结合的特征提取方法,同时提取到声学回声的时序特征和不同级别的抽象特征,且充分利用近端语音、近端麦克风语音和声学回声之间的幅度谱相似性的特点,引入它们之间的谱归一化互相关系数,构造了一种改进的理想二值掩蔽(iIBM)作为训练目标,此外引入深度可分离卷积使模型参数量减少了3.42 MB。实验结果表明:双端通话环境下所提出的方法相比参考算法取得了更高的客观评价得分。Aiming at the problem that the traditional acoustic echo cancellation(AEC)method is difficult to achieve fast convergence and dynamic adaptation in doubleended speech scenario,an AEC method combining long shortterm memory(LSTM)and residual neural network(ResNet)is proposed.By using the feature extraction method combining LSTM and ResNet,the temporal features of acoustic echo and abstract features of different levels are extracted at the same time,and the similarity of amplitude spectrum between proximal speech,proximal microphone speech and acoustic echo is fully utilized,and the spectral normalized cross correlation coefficient between them is introduced.An improved ideal binary mask(iIBM)is constructed as the training target,and the number of parameters of model is reduced by 3.42 MB by introducing depthwise separable convolution(DSC).The experimental results show that the proposed method achieves higher objective evaluation scores than the reference algorithm in the dualend communication environment.

关 键 词：声学回声消除 双端讲话场景 长短时记忆网络 残差神经网络 理想二值掩蔽 深度可分离卷积 

分 类 号：TN912[电子电信—通信与信息系统] TP391[电子电信—信息与通信工程]