低信噪比条件下深度学习麦克风阵列波束形成  

Deep learning microphone array beamforming under low signal noise ratio

作　　者：钟双莲 童峰[1,2,3] 刘雨佶 章宇栋 陈东升 ZHONG Shuanglian;TONG Feng;LIU Yuji;ZHANG Yudong;CHEN Dongsheng(Key Laboratory of Underwater Acoustic Communication and Marine Information Technology of the Ministry of Education,Xiamen University,Xiamen 361005,China;College of Ocean and Earth Sicences,Xiamen University,Xiamen 361102,China;Shenzhen Research Institute of Xiamen University,Shenzhen 518000,China)

机构地区：[1]厦门大学水声通信与海洋信息技术教育部重点实验室,福建厦门361005 [2]厦门大学海洋与地球学院,福建厦门361102 [3]厦门大学深圳研究院,广东深圳518000

出　　处：《兵器装备工程学报》2022年第11期273-277,296,共6页Journal of Ordnance Equipment Engineering

基　　金：国家自然科学基金项目(11274259);福建省自然科学基金计划项目(2018J05071);深圳虚拟大学园扶持经费研发机构建设项目(YFJGJS1.0)。

摘　　要：随着智能装备、可穿戴设备、智能家居、远程会议等语音交互领域对语音质量要求的不断提高,麦克风阵列被广泛用于前端语音增强,但低信噪比场景对传统麦阵增强方法造成了极大挑战。采用时频掩蔽深度学习麦阵算法,可利用语音和噪声频谱、通道间相位差特征估计掩蔽值实现增强,但其在低信噪比条件下性能无法保证。与声学模型联合训练的深度学习麦克风阵列算法,可以利用声学模型词错误率(word error rate,WER)等识别端反馈,以提高低信噪比条件下的识别性能,但该类方法无需重构信号,不输出增强语音信号,不适合远程会议、通话终端等需增强语音输出的应用。针对低信噪比条件,设计了一种基于空域代价函数的麦克风阵列波束形成深度学习网络,提高了低信噪比下的语音质量,并通过仿真和实验证明了该方法的有效性。With the increasing requirements for speech quality in the field of voice interaction,microphone arrays are widely used for front-end speech enhancement.However,low signal-to-noise ratio(SNR)scenarios pose great challenges to conventional approaches.The deep learning optimized time-frequency masking beamforming can achieve voice enhancement by extracting the spectral and inter-channel phase difference features,the performance of which degrades under low SNR scenarios.Deep learning microphone array algorithms trained jointly with acoustic models use the feedback such as word error rate(WER)to improve the recognition performance.However,such methods do not require signal reconstruction and do not output enhanced speech signals,which are not suitable for applications requiring enhanced speech output.A microphone array beamforming deep learning network based on the space domain cost function is designed to improve the speech quality under low SNR conditions,and the effectiveness of this method is verified by simulation and experiment.

关 键 词：波束形成 麦克风阵列 长短期记忆人工神经网络 

分 类 号：TN929.3[电子电信—通信与信息系统]