基于深度学习的听觉倒谱系数语音增强算法  

Speech enhancement based on auditory cepstral coefficient with deep learning

作　　者：李如玮 孙晓月 刘亚楠 李涛 LI Ruwei;SUN Xiaoyue;LIU Yanan;LI Tao(College of Information and Communications Engineering Faculty of Information Technology,Beijing University of Technology,Beijing 100124,China)

机构地区：[1]北京工业大学信息学部信息与通信工程学院

出　　处：《华中科技大学学报（自然科学版）》2019年第9期78-83,共6页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：北京市教委科技计划面上项目(KM201510005007);国家自然科学基金资助项目(51477028)

摘　　要：针对现有语音增强算法在低信噪比(SNR)非平稳噪声环境下的表现并不理想这一问题,提出了一种基于深度学习的语音增强算法.首先,构建了一个深度神经网络(DNN),然后从四个不同分辨率的耳蜗中提取了多分辨率听觉倒谱系数(MRACC)作为神经网络的输入,该系数既关注了细节的高分辨率特征,又把握了全局性的低分辨率特征;其次,跟踪噪声变化构建了一个自适应掩蔽阈值(AM)作为神经网络的训练目标,该阈值能够依据噪声调节理想二值掩蔽(IBM)和理想软掩蔽(IRM)的权重;最后,将估计的自适应掩蔽阈值用于对含噪语音进行增强.实验结果表明:相较于对比算法,该算法不仅可以进一步提高语音质量和可懂度,而且能够抑制更多的噪声.The performance of the existing speech enhancement algorithms is not ideal in low signal-to-noise ratio(SNR)non-stationary noise environments.In order to resolve this problem,a novel speech enhancement algorithm was presented.First,a fully connected deep neural network(DNN)was constructed,and a multi-resolution auditory cepstral coefficient(MRACC)was extracted from four cochleagrams of different resolutions as the input of neural network,which could capture the local information and spectrotemporal context.Second,an adaptive mask(AM)which can adjust the weight of ideal binary mask(IBM)and ideal ratio mask(IRM)according to noise change was put forward in this paper.Finally,the estimated AM was used to achieve the enhanced speech.The proposed algorithm shows that it not only further improves speech quality and intelligibility,but also suppresses more noise than the contrast algorithms by experimental results.

关 键 词：语音增强 深度神经网络 听觉倒谱系数 自适应掩蔽阈值 低信噪比 噪声跟踪 

分 类 号：TN912.35[电子电信—通信与信息系统] TP18[电子电信—信息与通信工程]