基于不确定性感知的语音分离方法  

Speech Separation Method Based on Uncertainty Perception

作　　者：涂斌炜 吕俊[1] Tu Binwei;Lü Jun(School of Automation,Guangdong University of Technology,Guangzhou 510006,China)

机构地区：[1]广东工业大学自动化学院,广东广州510006

出　　处：《自动化与信息工程》2021年第1期35-40,共6页Automation & Information Engineering

基　　金：广东省自然科学基金(2018A030313306)。

摘　　要：为抵御噪声的干扰,提出一种基于不确定性感知的语音分离方法。在训练阶段,采用双链路架构分别学习噪声和语音源成分的编解码子网和分离子网;在测试阶段,以闭式解的形式自适应更新噪声编码子网,减小训练与测试噪声在特征空间的均值偏移,降低认知不确定性,并尽量保持重要参数不变,间接限制语音分离的经验误差。在公开数据集LibriSpeech, NoiseX和NonSpeech上的实验结果表明:本文提出的方法能够快速有效地提高噪声干扰下语音分离的尺度不变信噪比。In order to resist the disturbances of noises, we proposed a speech separation method based on uncertainty perception. In the training phase, a two-link architecture is adopted to learn the codec subnet and separate subnet of noise and speech source components respectively. In the testing phase, the noise coding subnet is updated adaptively in the form of closed solution, so as to reduce the mean deviation of training and testing noises in the feature space, reduce cognitive uncertainty, keep the important parameters unchanged as far as possible, and indirectly limit the empirical error of speech separation. Experimental results on the public datasets LibriSpeech, NoiseX and NonSpeech show that the proposed approach can rapidly and effectively improve the scaleinvariant source-to-noise ratio of speech separation under the interferences of unknown noises.

关 键 词：语音分离 噪声干扰 不确定性感知 

分 类 号：TN912[电子电信—通信与信息系统]